Introduction and statements
The conjecture. Let R R R be the curvature tensor of a Riemannian manifold M M M . For a tangent vector x x x , the Jacobi operator J x y = R ( y , x ) x J_x y = R(y,x)x J x y = R ( y , x ) x is a symmetric operator on x ⊥ x^\perp x ⊥ . The manifold is pointwise Osserman if, at each point, the eigenvalues of J x J_x J x are the same for all unit vectors x x x , and globally Osserman if they are also the same at all points. Flat spaces and the rank-one symmetric spaces (spheres, the complex and quaternionic projective spaces, the Cayley projective plane, and their noncompact duals) are globally Osserman. Osserman asked in 1990 whether these are the only examples, locally [28 ] ; this is the Osserman [28 ] conjecture.
Chi proved the conjecture in dimensions not divisible by four and in dimension four [8 ] . Nikolayevsky proved it in all the remaining dimensions, and in its pointwise form for every $n \ne 2, 4, with a single exception: dimension sixteen, when a Jacobi eigenvalue has multiplicity seven or eight [22 , 23 ] , [24 Corollary 1.1] . He reduced this case, and the corresponding case of the conformal Osserman conjecture, to an algebraic statement, the Nikolayevsky (2009) conjecture [26 Theorem 1 and p. 679] . For surveys of Osserman manifolds see [13 , 15 ] . We prove the Nikolayevsky (2009) conjecture (Theorem A); hence the Osserman conjecture is true in every dimension (Corollary B and Remark 1.1). Table 1 compares our results with earlier ones.
Our results
We prove the Nikolayevsky (2009) conjecture (Theorem A), to which Nikolayevsky reduced the last open case of the Osserman [28 ] conjecture [26 ] .
Hence the Osserman [28 ] conjecture is true in every dimension, and the conformal Osserman conjecture in every dimension greater than four (Corollary B and Remark 1.1); before this paper the only open case of each was dimension sixteen with a Jacobi eigenvalue of multiplicity seven or eight [8 , 22 , 23 , 24 , 25 , 26 ] .
We classify the Osserman tensors on R 16 \mathbb{R}^{16} R 16 with two Jacobi eigenvalues, of multiplicities 7 and 8 (Theorem C ); Nikolayevsky announced the manifold form in 2006 [24 Theorem 1.2] , and no proof appeared [1 Section 4] .
Every Jacobi-proportional Osserman tensor on R 16 \mathbb{R}^{16} R 16 is Clifford or of Cayley type (Corollary D); Andrejić had reduced this question to the patterns (8, 7) and (7, 7, 1) [1 Theorem 5] .
Table 1. Our results and what was known before. Clifford tensors are defined under Notation below, and tensors of Cayley type after Theorem A.
Notation. Let V V V be a Euclidean vector space and R R R an algebraic curvature tensor on V V V . Its Jacobi operator is J x y = R ( y , x ) x \mathcal{J}_x y = R(y,x)x J x y = R ( y , x ) x . The tensor is Osserman if the spectrum of J x \mathcal{J}_x J x on x ⊥ x^\perp x ⊥ is independent of the unit vector x x x ; all multiplicities below refer to this reduced operator on x ⊥ x^\perp x ⊥ , as in [24 p. 379] , [26 proof of Lemma 4] , and the multiplicity pattern of R R R is the list of these multiplicities, such as (8, 7) or (7, 7, 1). Set
R 1 ( u , v ) w = ⟨ v , w ⟩ u − ⟨ u , w ⟩ v , Π ( x ) = ∥ x ∥ 2 I − x x T , R_1(u,v)w = \langle v,w\rangle u - \langle u,w\rangle v,\qquad\Pi(x) = \lVert x\rVert^2 I - xx^T, R 1 ( u , v ) w = ⟨ v , w ⟩ u − ⟨ u , w ⟩ v , Π ( x ) = ∥ x ∥ 2 I − x x T ,
so that J x ( R 1 ) = Π ( x ) \mathcal{J}_x(R_1) = \Pi(x) J x ( R 1 ) = Π ( x ) ; here u v T uv^T u v T denotes the map z ↦ ⟨ v , z ⟩ u z \mapsto\langle v,z\rangle u z ↦ ⟨ v , z ⟩ u . Our sign convention for J \mathcal{J} J is opposite to that of [23 , 24 , 26 ] ; we note the sign changes where their formulas are used. For an orthogonal map T T T we write ( T ∗ R ) ( u , v ) w = T − 1 R ( T u , T v ) T w (T^{*}R)(u,v)w = T^{-1}R(Tu,Tv)Tw ( T ∗ R ) ( u , v ) w = T − 1 R ( T u , T v ) T w . A Clifford structure , or Cl ( ν ) \operatorname{Cl}(\nu) Cl ( ν ) -structure, is a family of skew-symmetric endomorphisms J 1 , … , J ν J_{1},\ldots,J_{\nu} J 1 , … , J ν with J i J j + J j J i = − 2 δ i j I J_{i}J_{j}+J_{j}J_{i}=-2\delta_{ij}I J i J j + J j J i = − 2 δ ij I , and R R R is Clifford if J x = λ 0 Π ( x ) + ∑ i η i ( J i x ) ( J i x ) T J_{x}=\lambda_{0}\Pi(x)+\sum_{i}\eta_{i}(J_{i}x)(J_{i}x)^{T} J x = λ 0 Π ( x ) + ∑ i η i ( J i x ) ( J i x ) T for constants λ 0 , η i \lambda_{0},\eta_{i} λ 0 , η i . Such tensors were introduced by Gilkey [14 ] and by Gilkey, Swann and Vanhecke [16 ] ; see also [23 Definition 3] . A symmetric Clifford system is a family H 0 , … , H m H_{0},\ldots,H_{m} H 0 , … , H m of symmetric endomorphisms with H i H j + H j H i = 2 δ i j I H_{i}H_{j}+H_{j}H_{i}=2\delta_{ij}I H i H j + H j H i = 2 δ ij I . Finally R C a y R_{\mathrm{Cay}} R Cay is the curvature tensor of the Cayley projective plane, normalized to have sectional curvatures between 1 and 4.
Main results. The Nikolayevsky (2009) conjecture is Conjecture A of arXiv:0910.1645v1 (p. 2); the published version [26 p. 679] states it without a number.
Theorem A (The Nikolayevsky conjecture of 2009). Let R R R be an Osserman algebraic curvature tensor on R 16 \mathbb{R}^{16} R 16 one of whose reduced Jacobi eigenvalues has multiplicity 7 or 8. Then R R R is Clifford, or R R R has exactly two reduced eigenvalues and R = a R 1 + b T ∗ R C a y R=aR_{1}+bT^{*}R_{\mathrm{Cay}} R = a R 1 + b T ∗ R Cay for constants a , b a,b a , b and some T ∈ O ( 16 ) T\in O(16) T ∈ O ( 16 ) .
We say that R R R is of Cayley type if R = a R 1 + b T ∗ R C a y R=aR_{1}+bT^{*}R_{\mathrm{Cay}} R = a R 1 + b T ∗ R Cay as in Theorem A.
Corollary B. Let M M M be a connected smooth Riemannian manifold of dimension 16.
If M M M is pointwise Osserman, it is flat or locally isometric to a rank-one symmetric space.
If M M M is conformally Osserman (its Weyl tensor is Osserman at every point), it is locally conformally equivalent to a Euclidean space or to a rank-one symmetric space.
Proof. By [26 Theorem 1 and the paragraph after it, p. 679] , both statements hold once the curvature tensor, respectively the Weyl tensor, is Clifford or a combination of R 1 R_{1} R 1 and an orthogonal conjugate of R C a y R_{\mathrm{Cay}} R Cay at every point. (In arXiv:0910.1645v1 this is Theorem 1, “Assuming Conjecture A, . . .”; the dimension-sixteen argument ends in its Section 3.4.) At points where some reduced eigenvalue has multiplicity 7 or 8, this is Theorem A. At the other points, the largest multiplicity of a reduced eigenvalue is at least 7, by Adams’ theorem on vector fields on S 15 S^{15} S 15 [23 proof of Theorems 1 and 2, pp. 3–4 of arXiv:math/0204258v2] ; hence it is at least 9, and the tensor is Clifford by [24 Proposition 2.1] if it is 9 and by [23 Proposition 1] if it is at least 10, as summarized in [26 pp. 678 and 680] .
Remark 1.1. With [24 Corollary 1.1] , part (1) completes the proof of the Osserman conjecture: every globally Osserman manifold, and every pointwise Osserman manifold of dimension n ≠ 2 , 4 n\ne2,4 n = 2 , 4 , is flat or locally rank-one symmetric. With [25 Theorem 1] , part (2) proves the conformal Osserman conjecture in every dimension n > 4 n>4 n > 4 .
The proof starts with the case of two eigenvalues, which also fixes the normalization of R C a y R_{\mathrm{Cay}} R Cay .
Theorem C. Let R R R be an Osserman algebraic curvature tensor on R 16 \mathbb{R}^{16} R 16 with exactly two reduced eigenvalues, of multiplicities 7 and 8. There are constants μ \mu μ and κ ≠ 0 \kappa\ne0 κ = 0 such that J x = μ Π ( x ) + κ A ( x ) J_{x}=\mu\Pi(x)+\kappa A(x) J x = μ Π ( x ) + κ A ( x ) , where, in suitable orthogonal coordinates, A A A has one of the forms
A ( x ) = ∑ i = 1 7 ( J i x ) ( J i x ) T , for a Cl ( 7 ) -structure , (1) A(x)=\sum_{i=1}^{7}(J_{i}x)(J_{i}x)^{T}, \qquad\text{for a }\operatorname{Cl}(7)\text{-structure}, \tag*{(1)} A ( x ) = i = 1 ∑ 7 ( J i x ) ( J i x ) T , for a Cl ( 7 ) -structure , (1) A ( x ) = Π ( x ) − ∑ i = 1 8 ( J i x ) ( J i x ) T , for a Cl ( 8 ) -structure , (2) A(x)=\Pi(x)-\sum_{i=1}^{8}(J_{i}x)(J_{i}x)^{T}, \qquad\text{for a }\operatorname{Cl}(8)\text{-structure}, \tag*{(2)} A ( x ) = Π ( x ) − i = 1 ∑ 8 ( J i x ) ( J i x ) T , for a Cl ( 8 ) -structure , (2) A ( x ) = 1 2 ( ∥ x ∥ 2 I + ∑ i = 0 8 ⟨ H i x , x ⟩ H i ) − x x T , for a symmetric Clifford system H 0 , … , H 8 . (3) A(x)=\frac{1}{2}\left(\lVert x\rVert^{2}I+\sum_{i=0}^{8}\langle H_{i}x,x\rangle H_{i}\right)-xx^{T}, \qquad\text{for a symmetric Clifford system }H_{0},\ldots,H_{8}. \tag*{(3)} A ( x ) = 2 1 ( ∥ x ∥ 2 I + i = 0 ∑ 8 ⟨ H i x , x ⟩ H i ) − x x T , for a symmetric Clifford system H 0 , … , H 8 . (3)
The first two cases are Clifford. In the third, R = ( μ − κ / 3 ) R 1 + ( κ / 3 ) T ∗ R C a y \mathcal{R}=(\mu-\kappa/3)R_1+(\kappa/3)T^*R_{\mathrm{Cay}} R = ( μ − κ /3 ) R 1 + ( κ /3 ) T ∗ R Cay for some T ∈ O ( 16 ) T\in O(16) T ∈ O ( 16 ) .
We have not found this classification in the literature: [26 Lemma 4] takes it from Conjecture A, and [2 ] treats dimensions twice an odd number only. Corollary B(1) gives the theorem announced by Nikolayevsky in 2006 [24 Theorem 1.2] : a pointwise Osserman 16-manifold with two Jacobi eigenvalues of multiplicities 7 and 8 is locally isometric, up to scaling, to the Cayley projective plane or its noncompact dual. Indeed, such a manifold is flat or locally isometric to a rank-one symmetric space, and among these only the Cayley projective plane and its noncompact dual have Jacobi eigenvalues of multiplicities 7 and 8: a flat space or a space of constant curvature has one Jacobi eigenvalue, and C P 8 \mathbb{C}P^8 C P 8 , H P 4 \mathbb{H}P^4 H P 4 and their duals have multiplicities 1, 14 and 3, 12.
For an Osserman tensor, Jacobi-proportionality means
∥ x ∥ 2 ⟨ π i ( x ) y , y ⟩ = ∥ y ∥ 2 ⟨ π i ( y ) x , x ⟩ ( x , y ≠ 0 , every i ) , (4) \lVert x\rVert^2\langle\pi_i(x)y,y\rangle=\lVert y\rVert^2\langle\pi_i(y)x,x\rangle\quad(x,y\ne0,\ \text{every }i), \tag*{(4)} ∥ x ∥ 2 ⟨ π i ( x ) y , y ⟩ = ∥ y ∥ 2 ⟨ π i ( y ) x , x ⟩ ( x , y = 0 , every i ) , (4)
where, if the reduced eigenvalues are ε 1 ∥ x ∥ 2 , … , ε p ∥ x ∥ 2 \varepsilon_1\lVert x\rVert^2,\ldots,\varepsilon_p\lVert x\rVert^2 ε 1 ∥ x ∥ 2 , … , ε p ∥ x ∥ 2 , π i ( x ) \pi_i(x) π i ( x ) is the orthogonal projection onto the ε i ∥ x ∥ 2 \varepsilon_i\lVert x\rVert^2 ε i ∥ x ∥ 2 -eigenspace of J x J_x J x on x ⊥ x^\perp x ⊥ , regarded as an operator on R 16 \mathbb{R}^{16} R 16 that vanishes on R x \mathbb{R}x R x ; in particular, if 0 is a reduced eigenvalue, its eigenspace does not contain x x x .
Corollary D. Every Jacobi-proportional Riemannian Osserman algebraic curvature tensor on R 16 \mathbb{R}^{16} R 16 is Clifford or a linear combination of R 1 R_1 R 1 and an orthogonal conjugate of R C a y R_{\mathrm{Cay}} R Cay .
This settles the multiplicity patterns (8, 7) and (7, 7, 1), to which Andrejić reduced the Jacobi-proportional case [1 Theorem 5] , [3 Theorem 3] . It is proved in § 7 from Theorem C and Theorem E below, and it is used in the final step of the proof of Theorem A.
Theorem E. Let R R R be an Osserman algebraic curvature tensor on R 16 \mathbb{R}^{16} R 16 with exactly three reduced eigenvalues. If their multiplicities are (7, 7, 1), (8, 6, 1), (8, 5, 2) or (7, 6, 2), then R R R is Clifford.
Organization. Sections 2–6 prove Theorem C : the Jacobi family defines a Cartan–Münzner quartic on R 32 \mathbb{R}^{32} R 32 , and the classification of isoparametric hypersurfaces with four principal curvatures of multiplicities (7, 8) [9 , 11 , 21 ] recovers the eigenspaces. § 7 proves Theorem E and Corollary D with the polynomial eigenfield method of [24 ] . § 8 proves Theorem F, which settles fifteen more multiplicity patterns. In the remaining sixteen patterns, § 9 proves Theorem G, the case where the eigenvalue of multiplicity 7 or 8 is the smallest or the largest; otherwise we use a rational factorization of the Jacobi family, a coherent factor, whose poles are excluded, or shown to force a Clifford structure, case by case in Sections 10–26. Section 27 assembles the cases and proves Theorem A.
An expanded version of this paper, with many other results, is posted on Zenodo [17 ] .
Normalization and the quartic
Let R R R be as in Theorem C . Write the reduced eigenvalues as μ ∥ x ∥ 2 \mu\lVert x\rVert^2 μ ∥ x ∥ 2 of multiplicity 8 and ( μ + κ ) ∥ x ∥ 2 (\mu+\kappa)\lVert x\rVert^2 ( μ + κ ) ∥ x ∥ 2 of multiplicity 7, and put A ( x ) = ( J x − μ Π ( x ) ) / κ A(x)=(J_x-\mu\Pi(x))/\kappa A ( x ) = ( J x − μ Π ( x )) / κ . Then A A A is a homogeneous quadratic symmetric matrix family satisfying
A ( x ) x = 0 , A ( x ) 2 = ∥ x ∥ 2 A ( x ) , tr A ( x ) = 7 ∥ x ∥ 2 . (5) A(x)x=0,\qquad A(x)^2=\lVert x\rVert^2A(x),\qquad\operatorname{tr}A(x)=7\lVert x\rVert^2. \tag*{(5)} A ( x ) x = 0 , A ( x ) 2 = ∥ x ∥ 2 A ( x ) , tr A ( x ) = 7 ∥ x ∥ 2 . (5)
The Jacobi family determines its curvature tensor. In fact, for any symmetric quadratic family A A A with A ( x ) x = 0 A(x)x=0 A ( x ) x = 0 , let B ( u , v ) = ( A ( u + v ) − A ( u ) − A ( v ) ) / 2 B(u,v)=(A(u+v)-A(u)-A(v))/2 B ( u , v ) = ( A ( u + v ) − A ( u ) − A ( v )) /2 . Polarization gives B ( x , y ) x = − A ( x ) y / 2 B(x,y)x=-A(x)y/2 B ( x , y ) x = − A ( x ) y /2 . Since B ( x , y ) B(x,y) B ( x , y ) is symmetric,
⟨ A ( x ) y , y ⟩ = − 2 ⟨ B ( x , y ) x , y ⟩ = − 2 ⟨ B ( y , x ) y , x ⟩ = ⟨ A ( y ) x , x ⟩ . \langle A(x)y,y\rangle=-2\langle B(x,y)x,y\rangle=-2\langle B(y,x)y,x\rangle=\langle A(y)x,x\rangle. ⟨ A ( x ) y , y ⟩ = − 2 ⟨ B ( x , y ) x , y ⟩ = − 2 ⟨ B ( y , x ) y , x ⟩ = ⟨ A ( y ) x , x ⟩ .
Jacobi-family reconstruction [3 Theorem 1] therefore applies; specializing its defining formula [3 (12), p. 8468] to a quadratic family gives the curvature tensor
R A ( u , v ) w = 2 3 ( B ( v , w ) u − B ( u , w ) v ) (6) R_A(u,v)w = \frac{2}{3}\bigl(\mathcal{B}(v,w)u-\mathcal{B}(u,w)v\bigr) \tag*{(6)} R A ( u , v ) w = 3 2 ( B ( v , w ) u − B ( u , w ) v ) (6)
with Jacobi family A A A . Conversely, if R R R is an algebraic curvature tensor with Jacobi family A A A , then B ( v , w ) u = 1 2 ( R ( u , v ) w + R ( u , w ) v ) \mathcal{B}(v,w)u = \frac{1}{2}\bigl(R(u,v)w+R(u,w)v\bigr) B ( v , w ) u = 2 1 ( R ( u , v ) w + R ( u , w ) v ) , and the first Bianchi identity turns (6) into R A = R R_A=R R A = R ; so the Jacobi family determines the tensor.
Set
Q ( x ) = A ( x ) + x x T , L ( x ) = R x + Im A ( x ) . Q(x)=A(x)+xx^{T},\qquad L(x)=\mathbb{R}x+\operatorname{Im}A(x). Q ( x ) = A ( x ) + x x T , L ( x ) = R x + Im A ( x ) .
For x ≠ 0 x\ne0 x = 0 , Q ( x ) = ∥ x ∥ 2 π L ( x ) Q(x)=\lVert x\rVert^2\pi_{L(x)} Q ( x ) = ∥ x ∥ 2 π L ( x ) and dim L ( x ) = 8 \dim L(x)=8 dim L ( x ) = 8 . In particular,
Q ( x ) 2 = ∥ x ∥ 2 Q ( x ) , Q ( x ) x = ∥ x ∥ 2 x , tr Q ( x ) = 8 ∥ x ∥ 2 . (7) Q(x)^2=\lVert x\rVert^2Q(x),\qquad Q(x)x=\lVert x\rVert^2x,\qquad\operatorname{tr}Q(x)=8\lVert x\rVert^2. \tag*{(7)} Q ( x ) 2 = ∥ x ∥ 2 Q ( x ) , Q ( x ) x = ∥ x ∥ 2 x , tr Q ( x ) = 8 ∥ x ∥ 2 . (7)
The symmetry ⟨ A ( x ) y , y ⟩ = ⟨ A ( y ) x , x ⟩ \langle A(x)y,y\rangle=\langle A(y)x,x\rangle ⟨ A ( x ) y , y ⟩ = ⟨ A ( y ) x , x ⟩ shown above implies that the biquadratic form
b ( x , y ) = ⟨ Q ( x ) y , y ⟩ = ⟨ A ( x ) y , y ⟩ + ⟨ x , y ⟩ 2 (8) b(x,y)=\langle Q(x)y,y\rangle=\langle A(x)y,y\rangle+\langle x,y\rangle^2 \tag*{(8)} b ( x , y ) = ⟨ Q ( x ) y , y ⟩ = ⟨ A ( x ) y , y ⟩ + ⟨ x , y ⟩ 2 (8)
is symmetric in x , y x,y x , y . It vanishes when x = 0 x=0 x = 0 or y = 0 y=0 y = 0 and satisfies b ( x , x ) = ∥ x ∥ 4 b(x,x)=\lVert x\rVert^4 b ( x , x ) = ∥ x ∥ 4 .
On R 32 = R 16 ⊕ R 16 \mathbb{R}^{32}=\mathbb{R}^{16}\oplus\mathbb{R}^{16} R 32 = R 16 ⊕ R 16 write z = ( x , y ) z=(x,y) z = ( x , y ) , s = ∥ x ∥ 2 + ∥ y ∥ 2 s=\lVert x\rVert^2+\lVert y\rVert^2 s = ∥ x ∥ 2 + ∥ y ∥ 2 , and define
F ( z ) = s 2 − 8 b ( x , y ) . (9) F(z)=s^2-8b(x,y). \tag*{(9)} F ( z ) = s 2 − 8 b ( x , y ) . (9)
Lemma 2.1. The polynomial F F F satisfies
∥ ∇ F ∥ 2 = 16 s 3 , Δ F = 8 s . \lVert\nabla F\rVert^2=16s^3,\qquad\Delta F=8s. ∥ ∇ F ∥ 2 = 16 s 3 , Δ F = 8 s . Its restriction to S 31 S^{31} S 31 takes both values 1 1 1 and − 1 -1 − 1 .
Proof. Symmetry of b b b gives ∇ y b = 2 Q ( x ) y \nabla_y b=2Q(x)y ∇ y b = 2 Q ( x ) y and ∇ x b = 2 Q ( y ) x \nabla_x b=2Q(y)x ∇ x b = 2 Q ( y ) x . By (7) ,
∥ ∇ b ∥ 2 = 4 s b , ⟨ z , ∇ b ⟩ = 4 b , Δ b = 2 tr Q ( x ) + 2 tr Q ( y ) = 16 s . \lVert\nabla b\rVert^2=4sb,\qquad\langle z,\nabla b\rangle=4b,\qquad\Delta b=2\operatorname{tr}Q(x)+2\operatorname{tr}Q(y)=16s. ∥ ∇ b ∥ 2 = 4 s b , ⟨ z , ∇ b ⟩ = 4 b , Δ b = 2 tr Q ( x ) + 2 tr Q ( y ) = 16 s . Thus
∥ ∇ F ∥ 2 = 16 s 3 − 64 s ⟨ z , ∇ b ⟩ + 64 ∥ ∇ b ∥ 2 = 16 s 3 . \lVert\nabla F\rVert^2=16s^3-64s\langle z,\nabla b\rangle+64\lVert\nabla b\rVert^2=16s^3. ∥ ∇ F ∥ 2 = 16 s 3 − 64 s ⟨ z , ∇ b ⟩ + 64 ∥ ∇ b ∥ 2 = 16 s 3 . Since Δ s 2 = 4 ( 32 + 2 ) s = 136 s \Delta s^2=4(32+2)s=136s Δ s 2 = 4 ( 32 + 2 ) s = 136 s , we obtain Δ F = 8 s \Delta F=8s Δ F = 8 s . For unit x x x , F ( x , 0 ) = 1 F(x,0)=1 F ( x , 0 ) = 1 . If x , y x,y x , y are unit and y ∈ L ( x ) y\in L(x) y ∈ L ( x ) , then F ( ( x , y ) / 2 ) = 1 − 8 / 4 = − 1 F((x,y)/\sqrt{2})=1-8/4=-1 F (( x , y ) / 2 ) = 1 − 8/4 = − 1 .
The classification of isoparametric hypersurfaces
We record the external results that are used.
Münzner. For a homogeneous polynomial of degree g g g satisfying ∥ ∇ F ∥ 2 = g 2 ∥ z ∥ 2 g − 2 \lVert\nabla F\rVert^2=g^2\lVert z\rVert^{2g-2} ∥ ∇ F ∥ 2 = g 2 ∥ z ∥ 2 g − 2 and Δ F = c ∥ z ∥ g − 2 \Delta F=c\lVert z\rVert^{g-2} Δ F = c ∥ z ∥ g − 2 on R n + 1 \mathbb{R}^{n+1} R n + 1 , nonconstant on the sphere, [21 Satz 3, p. 65] shows that 0 0 0 is a regular value of F ∣ S n F|S^n F ∣ S n and that the zero level is an isoparametric hypersurface; by its proof (p. 66) the same holds for every level in ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) . Its Folgerung (corollary) on p. 66 identifies F F F with the Cartan polynomial of these hypersurfaces when both g ( n − 1 ) − c g(n-1)-c g ( n − 1 ) − c and g ( n − 1 ) + c g(n-1)+c g ( n − 1 ) + c are nonzero. Satz 4(a), p. 67, gives connected level sets. By Lemma 2.1 these hypotheses hold with g = 4 g=4 g = 4 , n = 31 n=31 n = 31 and c = 8 c=8 c = 8 , whereas g ( n − 1 ) = 120 g(n-1)=120 g ( n − 1 ) = 120 . The resulting hypersurfaces have four principal curvatures, and Münzner’s relations g ( m 1 + m 2 ) = 2 ( n − 1 ) g(m_1+m_2)=2(n-1) g ( m 1 + m 2 ) = 2 ( n − 1 ) and c = g 2 ( m 2 − m 1 ) / 2 c=g^2(m_2-m_1)/2 c = g 2 ( m 2 − m 1 ) /2 [21 Satz 2, p. 63] give
m 1 + m 2 = 15 , 8 ( m 2 − m 1 ) = 8 , m_1+m_2=15,\qquad8(m_2-m_1)=8, m 1 + m 2 = 15 , 8 ( m 2 − m 1 ) = 8 ,
so their multiplicities are ( 7 , 8 ) (7,8) ( 7 , 8 ) .
Chi and Ferus–Karcher–Münzner. Chi’s classification [9 abstract, p. 225] gives exactly three families with unordered multiplicities { 7 , 8 } \{7,8\} { 7 , 8 } . They are the FKM family with parameter m = 7 m=7 m = 7 and the two FKM families with m = 8 m=8 m = 8 . The first is the Ozeki–Takeuchi family, as identified in [11 Section 5.9, p. 19] (page numbers for FKM refer to its English translation, arXiv:1112.2780); the two ( 8 , 7 ) (8,7) ( 8 , 7 ) families are distinguished in its Section 6.6 [11 p. 22] . For a symmetric Clifford system P 0 , … , P m P_0,\ldots,P_m P 0 , … , P m on R 32 \mathbb{R}^{32} R 32 with m ≥ 1 m\ge1 m ≥ 1 , [11, Theorem 4.1, p. 8; Section 4.3, p. 11] gives
F P ( z ) = ∥ z ∥ 4 − 2 ∑ i = 0 m ⟨ P i z , z ⟩ 2 , ( m 1 , m 2 ) = ( m , 15 − m ) . (10) F_P(z)=\lVert z\rVert^4-2\sum_{i=0}^{m}\langle P_i z,z\rangle^2,\qquad(m_1,m_2)=(m,15-m). \tag*{(10)} F P ( z ) = ∥ z ∥ 4 − 2 i = 0 ∑ m ⟨ P i z , z ⟩ 2 , ( m 1 , m 2 ) = ( m , 15 − m ) . (10)
By the same Folgerung of Münzner, with c = ± 8 ≠ ± 120 c=\pm8\ne\pm120 c = ± 8 = ± 120 , each F P F_P F P used here is the Cartan polynomial of its zero level.
We need congruence of the polynomials, not only of a level hypersurface. The zero level of F F F has principal curvatures cot ( π / 8 + j π / 4 ) \cot(\pi/8+j\pi/4) cot ( π /8 + j π /4 ) , 0 ≤ j ≤ 3 0\le j\le3 0 ≤ j ≤ 3 . A congruent member of the FKM family must therefore be its zero member, with one of the two choices of normal. Let U ∈ O ( 32 ) U\in O(32) U ∈ O ( 32 ) map the zero level of F F F onto the zero level of F P F_P F P . Both zero levels are connected [21 Satz 4(a), p. 67] . By [21 Folgerung (iii), p. 66] , F F F and F P ∘ U F_P\circ U F P ∘ U are both Cartan polynomials of this hypersurface. The Cartan polynomial of a connected oriented isoparametric hypersurface is determined by it, and reversing the orientation replaces it by its negative [21 §4, p. 65] ; explicitly F ( r q ) = r 4 cos 4 t ( q ) F(rq)=r^4\cos4t(q) F ( r q ) = r 4 cos 4 t ( q ) , where t ( q ) ∈ [ 0 , π / 4 ] t(q)\in[0,\pi/4] t ( q ) ∈ [ 0 , π /4 ] is the spherical distance from q ∈ S 31 q\in S^{31} q ∈ S 31 to the focal submanifold { F = 1 } ∩ S 31 \{F=1\}\cap S^{31} { F = 1 } ∩ S 31 [21, proof of Satz 3, p. 66; Satz 4(c), p. 67]. Hence F = ε F P ∘ U F=\varepsilon F_P\circ U F = ε F P ∘ U with ε ∈ { 1 , − 1 } \varepsilon\in\{1,-1\} ε ∈ { 1 , − 1 } .
Absorb U U U into the matrices P i P_i P i . Since Δ F P = ( 136 − 16 ( m + 1 ) ) s \Delta F_P=(136-16(m+1))s Δ F P = ( 136 − 16 ( m + 1 )) s , comparison with Δ F = 8 s \Delta F=8s Δ F = 8 s gives precisely
4 b ( x , y ) = ∑ i = 0 7 ⟨ P i z , z ⟩ 2 (11) 4b(x,y)=\sum_{i=0}^{7}\langle P_i z,z\rangle^2 \tag*{(11)} 4 b ( x , y ) = i = 0 ∑ 7 ⟨ P i z , z ⟩ 2 (11)
or
4 b ( x , y ) = s 2 − ∑ i = 0 8 ⟨ P i z , z ⟩ 2 . (12) 4b(x,y)=s^2-\sum_{i=0}^{8}\langle P_i z,z\rangle^2. \tag*{(12)} 4 b ( x , y ) = s 2 − i = 0 ∑ 8 ⟨ P i z , z ⟩ 2 . (12)
These identities hold in the original coordinates ( x , y ) (x,y) ( x , y ) ; no preservation of the two coordinate summands by U U U has been assumed.
A lemma on quadratic maps between spheres
Quadratic maps between spheres have a classical literature [30 , 31 ] . We give the short argument needed here. Part (b) below is the quadratic case of [31 p. 163] ; we include its short proof.
Lemma 4.1. Let H 0 , … , H m H_0,\ldots,H_m H 0 , … , H m be symmetric n × n n\times n n × n matrices and h ( x ) = ( ⟨ H i x , x ⟩ ) i = 0 m h(x)=(\langle H_i x,x\rangle)_{i=0}^{m} h ( x ) = (⟨ H i x , x ⟩ ) i = 0 m . Suppose ∥ h ( x ) ∥ = ∥ x ∥ 2 \lVert h(x)\rVert=\lVert x\rVert^2 ∥ h ( x )∥ = ∥ x ∥ 2 . For a ∈ R m + 1 a\in\mathbb{R}^{m+1} a ∈ R m + 1 put H ( a ) = ∑ i a i H i H(a)=\sum_i a_iH_i H ( a ) = ∑ i a i H i .
(a) If a ∈ h ( S n − 1 ) a\in h(S^{n-1}) a ∈ h ( S n − 1 ) , then I − H ( a ) ≥ 0 I-H(a)\ge0 I − H ( a ) ≥ 0 and the fibre over a a a is the unit sphere of U a = ker ( I − H ( a ) ) U_a=\ker(I-H(a)) U a = ker ( I − H ( a )) , where dim U a ≥ n − m \dim U_a\ge n-m dim U a ≥ n − m .
(b) If n > 2 m n>2m n > 2 m , then h h h is constant on S n − 1 S^{n-1} S n − 1 .
(c) If n = 2 m n=2m n = 2 m , m ≥ 1 m\ge1 m ≥ 1 , and h h h is not constant on S n − 1 S^{n-1} S n − 1 , then H i H j + H j H i = 2 δ i j I H_iH_j+H_jH_i=2\delta_{ij}I H i H j + H j H i = 2 δ ij I .
Proof. For unit a a a , x x x , ⟨ H ( a ) x , x ⟩ = ⟨ a , h ( x ) ⟩ ≤ 1 \langle H(a)x,x\rangle=\langle a,h(x)\rangle\leq1 ⟨ H ( a ) x , x ⟩ = ⟨ a , h ( x )⟩ ≤ 1 , with equality exactly when h ( x ) = a h(x)=a h ( x ) = a . This proves the fibre description. Fix a unit x x x in that fibre and rotate the index space so that a = e 0 a=e_0 a = e 0 . Then H 0 x = x H_0x=x H 0 x = x . Comparing the t 2 t^2 t 2 coefficients of ∥ h ( x + t y ) ∥ 2 = ∥ x + t y ∥ 4 \lVert h(x+ty)\rVert^2=\lVert x+ty\rVert^4 ∥ h ( x + t y ) ∥ 2 = ∥ x + t y ∥ 4 gives
2 ⟨ H 0 y , y ⟩ + 4 ∑ i = 0 m ⟨ H i x , y ⟩ 2 = 2 ∥ y ∥ 2 + 4 ⟨ x , y ⟩ 2 2\langle H_0y,y\rangle+4\sum_{i=0}^{m}\langle H_ix,y\rangle^2=2\lVert y\rVert^2+4\langle x,y\rangle^2 2 ⟨ H 0 y , y ⟩ + 4 i = 0 ∑ m ⟨ H i x , y ⟩ 2 = 2 ∥ y ∥ 2 + 4 ⟨ x , y ⟩ 2
The i = 0 i=0 i = 0 terms cancel, so
I − H 0 = 2 ∑ i = 1 m ( H i x ) ( H i x ) T . (13) I-H_0=2\sum_{i=1}^{m}(H_ix)(H_ix)^{T}. \tag*{(13)} I − H 0 = 2 i = 1 ∑ m ( H i x ) ( H i x ) T . (13)
Its rank is at most m m m , proving (a).
The subspaces U a , U b U_a,U_b U a , U b for distinct values a ≠ b a\ne b a = b meet only at zero. If n > 2 m n>2m n > 2 m , two such subspaces would have dimensions summing to more than n n n . This proves (b).
Suppose n = 2 m n=2m n = 2 m and there is more than one value. For every a a a , comparison with another fibre gives dim U a = m \dim U_a=m dim U a = m . Hence the right side of (13) has rank m m m , and H 1 x , … , H m x H_1x,\ldots,H_mx H 1 x , … , H m x are independent and perpendicular to x x x . On T x S n − 1 = x ⊥ T_xS^{n-1}=x^\perp T x S n − 1 = x ⊥ ,
( d h ) x ( y ) = 2 ( ⟨ H i x , y ⟩ ) i = 0 m (dh)_x(y)=2\bigl(\langle H_ix,y\rangle\bigr)_{i=0}^{m} ( d h ) x ( y ) = 2 ( ⟨ H i x , y ⟩ ) i = 0 m
has rank m m m . The image of h h h is open in S m S^m S m and is compact; since m ≥ 1 m\geq1 m ≥ 1 , it is all of S m S^m S m . In particular a a a and − a -a − a are values. Their fibres are the + 1 +1 + 1 and − 1 -1 − 1 eigenspaces of H ( a ) H(a) H ( a ) and each has dimension m m m . Thus H ( a ) 2 = I H(a)^2=I H ( a ) 2 = I for every unit a a a . Polarization in a a a proves (c). □
Recovering the original eigenspaces
Proposition 5.1. Identity (3.2) implies (1.1).
Proof. Setting y = 0 y=0 y = 0 in (3.2) shows that the upper-left diagonal block of each P i P_i P i is zero; setting x = 0 x=0 x = 0 gives the same for the lower-right block. Hence
P i = ( 0 C i T C i 0 ) , Q ( x ) = ∑ i = 0 7 ( C i x ) ( C i x ) T . P_i=\begin{pmatrix}0&C_i^T\\ C_i&0\end{pmatrix},\qquad Q(x)=\sum_{i=0}^{7}(C_ix)(C_ix)^T. P i = ( 0 C i C i T 0 ) , Q ( x ) = i = 0 ∑ 7 ( C i x ) ( C i x ) T . The Clifford relations give
C i T C j + C j T C i = 2 δ i j I , C i C j T + C j C i T = 2 δ i j I . C_i^TC_j+C_j^TC_i=2\delta_{ij}I,\qquad C_iC_j^T+C_jC_i^T=2\delta_{ij}I. C i T C j + C j T C i = 2 δ ij I , C i C j T + C j C i T = 2 δ ij I . Moreover Q ( x ) x = ∥ x ∥ 2 x Q(x)x=\lVert x\rVert^2x Q ( x ) x = ∥ x ∥ 2 x implies ∑ i ⟨ C i x , x ⟩ 2 = ∥ x ∥ 4 \sum_i\langle C_ix,x\rangle^2=\lVert x\rVert^4 ∑ i ⟨ C i x , x ⟩ 2 = ∥ x ∥ 4 . Apply Lemma 4.1(b) to the symmetric parts of C i C_i C i , with n = 16 n=16 n = 16 and m = 7 m=7 m = 7 . There is a unit w ∈ R 8 w\in\mathbb{R}^8 w ∈ R 8 such that ( ⟨ C i x , x ⟩ ) i = ∥ x ∥ 2 w (\langle C_ix,x\rangle)_i=\lVert x\rVert^2w (⟨ C i x , x ⟩ ) i = ∥ x ∥ 2 w . Rotating the index space gives w = e 0 w=e_0 w = e 0 . Since C 0 C_0 C 0 is orthogonal and ⟨ C 0 x , x ⟩ = ∥ x ∥ 2 \langle C_0x,x\rangle=\lVert x\rVert^2 ⟨ C 0 x , x ⟩ = ∥ x ∥ 2 , we have C 0 = I C_0=I C 0 = I . The remaining C i C_i C i are skew, and their relations become C i C j + C j C i = − 2 δ i j I C_iC_j+C_jC_i=-2\delta_{ij}I C i C j + C j C i = − 2 δ ij I . Subtracting x x T xx^T x x T from Q ( x ) Q(x) Q ( x ) proves the claim. □
Proposition 5.2. Identity (3.3) implies (1.2) or (1.3).
Proof. In the original splitting write P i = ( H i B i T B i D i ) P_i=\begin{pmatrix}H_i&B_i^T\\ B_i&D_i\end{pmatrix} P i = ( H i B i B i T D i ) . Setting y = 0 y=0 y = 0 and then x = 0 x=0 x = 0 in (3.3) gives
∑ i = 0 8 ⟨ H i x , x ⟩ 2 = ∥ x ∥ 4 , ∑ i = 0 8 ⟨ D i y , y ⟩ 2 = ∥ y ∥ 4 . \sum_{i=0}^{8}\langle H_ix,x\rangle^2=\lVert x\rVert^4,\qquad\sum_{i=0}^{8}\langle D_iy,y\rangle^2=\lVert y\rVert^4. i = 0 ∑ 8 ⟨ H i x , x ⟩ 2 = ∥ x ∥ 4 , i = 0 ∑ 8 ⟨ D i y , y ⟩ 2 = ∥ y ∥ 4 .
Write h ( x ) = ( ⟨ H i x , x ⟩ ) i h(x) = (\langle H_i x, x\rangle)_i h ( x ) = (⟨ H i x , x ⟩ ) i and d ( y ) = ( ⟨ D i y , y ⟩ ) i d(y) = (\langle D_i y, y\rangle)_i d ( y ) = (⟨ D i y , y ⟩ ) i .
If h h h is constant on the unit sphere, rotate the index so that H 0 = I H_0 = I H 0 = I , H i = 0 H_i = 0 H i = 0 for i ≥ 1 i \ge1 i ≥ 1 . The upper block of P 0 2 = I P_0^2 = I P 0 2 = I gives B 0 = 0 B_0 = 0 B 0 = 0 . Also tr P 0 = 0 \operatorname{tr} P_0 = 0 tr P 0 = 0 , because P 0 = − P 1 P 0 P 1 P_0 = -P_1P_0P_1 P 0 = − P 1 P 0 P 1 . Thus D 0 2 = I D_0^2 = I D 0 2 = I and tr D 0 = − 16 \operatorname{tr} D_0 = -16 tr D 0 = − 16 , forcing D 0 = − I D_0 = -I D 0 = − I . Anticommutation with P 0 = diag ( I , − I ) P_0 = \operatorname{diag}(I,-I) P 0 = diag ( I , − I ) makes the other eight matrices off-diagonal. Calling their lower blocks C 1 , … , C 8 C_1,\ldots,C_8 C 1 , … , C 8 and expanding (3.3) yields
Q ( x ) = ∥ x ∥ 2 I − ∑ i = 1 8 ( C i x ) ( C i x ) T . Q(x) = \lVert x\rVert^2 I - \sum_{i=1}^{8}(C_i x)(C_i x)^{T}. Q ( x ) = ∥ x ∥ 2 I − i = 1 ∑ 8 ( C i x ) ( C i x ) T .
The identity Q ( x ) x = ∥ x ∥ 2 x Q(x)x = \lVert x\rVert^2 x Q ( x ) x = ∥ x ∥ 2 x gives ∑ i ⟨ C i x , x ⟩ 2 = 0 \sum_i \langle C_i x,x\rangle^2 = 0 ∑ i ⟨ C i x , x ⟩ 2 = 0 . Each C i C_i C i is therefore skew, and the Clifford relations become those of a Cl ( 8 ) \operatorname{Cl}(8) Cl ( 8 ) -structure. This is (1.2).
If h h h is not constant, Lemma 4.1(c) makes the H i H_i H i a symmetric Clifford system. The equation H i 2 + B i T B i = I H_i^2 + B_i^{T}B_i = I H i 2 + B i T B i = I forces B i = 0 B_i = 0 B i = 0 . Expanding (3.3) now gives
b ( x , y ) = 1 2 ( ∥ x ∥ 2 ∥ y ∥ 2 − ⟨ h ( x ) , d ( y ) ⟩ ) . b(x,y) = \frac{1}{2}\left(\lVert x\rVert^2\lVert y\rVert^2 - \langle h(x),d(y)\rangle\right). b ( x , y ) = 2 1 ( ∥ x ∥ 2 ∥ y ∥ 2 − ⟨ h ( x ) , d ( y )⟩ ) .
Since b ( x , x ) = ∥ x ∥ 4 b(x,x) = \lVert x\rVert^4 b ( x , x ) = ∥ x ∥ 4 and ∥ h ( x ) ∥ = ∥ d ( x ) ∥ = ∥ x ∥ 2 \lVert h(x)\rVert= \lVert d(x)\rVert= \lVert x\rVert^2 ∥ h ( x )∥ = ∥ d ( x )∥ = ∥ x ∥ 2 , equality in Cauchy–Schwarz forces d ( x ) = − h ( x ) d(x) = -h(x) d ( x ) = − h ( x ) . Polarization gives D i = − H i D_i = -H_i D i = − H i . Consequently
Q ( x ) = 1 2 ( ∥ x ∥ 2 I + ∑ i = 0 8 ⟨ H i x , x ⟩ H i ) . Q(x) = \frac{1}{2}\left(\lVert x\rVert^2 I + \sum_{i=0}^{8}\langle H_i x,x\rangle H_i\right). Q ( x ) = 2 1 ( ∥ x ∥ 2 I + i = 0 ∑ 8 ⟨ H i x , x ⟩ H i ) .
which proves (1.3).
Identification of the Cayley tensor
We first prove two identities for an arbitrary symmetric Clifford system S 0 , … , S 8 S_0,\ldots,S_8 S 0 , … , S 8 on R 16 \mathbb{R}^{16} R 16 : (6.1) below and the norm identity ∑ i = 0 8 ⟨ S i x , x ⟩ 2 = ∥ x ∥ 4 \sum_{i=0}^{8}\langle S_i x,x\rangle^2 = \lVert x\rVert^4 ∑ i = 0 8 ⟨ S i x , x ⟩ 2 = ∥ x ∥ 4 . For q = ∥ x ∥ 2 q = \lVert x\rVert^2 q = ∥ x ∥ 2 put h i = ⟨ S i x , x ⟩ h_i = \langle S_i x,x\rangle h i = ⟨ S i x , x ⟩ and S h = ∑ i h i S i S_h = \sum_i h_i S_i S h = ∑ i h i S i . The products of zero, one, and four distinct generators are symmetric (a product of k k k distinct generators is ( − 1 ) k ( k − 1 ) / 2 (-1)^{k(k-1)/2} ( − 1 ) k ( k − 1 ) /2 times its transpose) and trace-orthogonal (the product of two distinct such monomials is, up to sign, a product of j j j distinct generators with 1 ≤ j ≤ 8 1 \le j \le8 1 ≤ j ≤ 8 , which anticommutes with a generator outside it if j j j is odd and with one inside it if j j j is even, so has trace zero), and their number is 1 + 9 + 126 = 136 1 + 9 + 126 = 136 1 + 9 + 126 = 136 . They therefore form a basis of the symmetric matrices. If L k \mathcal{L}_k L k is their span for degree k k k , the operator T ( Y ) = ∑ i S i Y S i T(Y) = \sum_i S_i Y S_i T ( Y ) = ∑ i S i Y S i acts on L k \mathcal{L}_k L k by ( − 1 ) k ( 9 − 2 k ) (-1)^k(9-2k) ( − 1 ) k ( 9 − 2 k ) , namely by 9 , − 7 , 1 9,-7,1 9 , − 7 , 1 for k = 0 , 1 , 4 k=0,1,4 k = 0 , 1 , 4 . These assertions also appear in [26 Lemma 3, pp. 6–7 of arXiv:0910.1645v1] . Expanding x x T = q I / 16 + S h / 16 + Y 4 xx^T = qI/16 + S_h/16 + Y_4 x x T = q I /16 + S h /16 + Y 4 therefore gives
Z x : = ∑ i ( S i x ) ( S i x ) T = x x T + 1 2 ( q I − S h ) . (14) Z_x := \sum_i (S_i x)(S_i x)^T = xx^T + \frac{1}{2}(qI-S_h). \tag*{(14)} Z x := i ∑ ( S i x ) ( S i x ) T = x x T + 2 1 ( q I − S h ) . (14)
Contraction with x x x yields ∥ h ∥ 2 = q 2 + ( q 2 − ∥ h ∥ 2 ) / 2 \lVert h\rVert^2 = q^2 + (q^2-\lVert h\rVert^2)/2 ∥ h ∥ 2 = q 2 + ( q 2 − ∥ h ∥ 2 ) /2 , so ∥ h ∥ = q \lVert h\rVert= q ∥ h ∥ = q .
For this system define Q ( x ) = ( q I + S h ) / 2 Q(x) = (qI+S_h)/2 Q ( x ) = ( q I + S h ) /2 , so that A ( x ) = Q ( x ) − x x T A(x) = Q(x)-xx^T A ( x ) = Q ( x ) − x x T is the matrix in (1.3) with H i = S i H_i = S_i H i = S i . For x ≠ 0 x \ne0 x = 0 set w = h / q w = h/q w = h / q . Since S h = q S w S_h = qS_w S h = q S w , we have Q ( x ) = q ( I + S w ) / 2 Q(x) = q(I+S_w)/2 Q ( x ) = q ( I + S w ) /2 . Also S w x = x S_wx = x S w x = x , since S w S_w S w is an orthogonal involution and ⟨ S w x , x ⟩ = q \langle S_wx,x\rangle= q ⟨ S w x , x ⟩ = q . Its two eigenspaces have dimension eight. For u ⊥ w u \perp w u ⊥ w , S u x S_u x S u x lies in its negative eigenspace. With (6.1) and S h = q S w S_h = qS_w S h = q S w ,
Q ( x ) = q 2 ( I + S w ) , Z x = x x T + q I − Q ( x ) , S h = 2 Q ( x ) − q I . Q(x) = \frac{q}{2}(I+S_w), \qquad Z_x = xx^T + qI-Q(x), \qquad S_h = 2Q(x)-qI. Q ( x ) = 2 q ( I + S w ) , Z x = x x T + q I − Q ( x ) , S h = 2 Q ( x ) − q I .
In octonionic coordinates x = ( x 1 , x 2 ) x=(x_1,x_2) x = ( x 1 , x 2 ) the standard system is
S 0 ( x 1 , x 2 ) = ( x 1 , − x 2 ) , S i ( x 1 , x 2 ) = ( e i x ‾ 2 , x ‾ 1 e i ) , 1 ≤ i ≤ 8. S_0(x_1,x_2)=(x_1,-x_2),\qquad S_i(x_1,x_2)=(e_i\overline{x}_2,\overline{x}_1e_i),\qquad1\le i\le8. S 0 ( x 1 , x 2 ) = ( x 1 , − x 2 ) , S i ( x 1 , x 2 ) = ( e i x 2 , x 1 e i ) , 1 ≤ i ≤ 8.
where ( e i ) i = 1 8 (e_i)_{i=1}^{8} ( e i ) i = 1 8 is an orthonormal basis of O \mathbb{O} O . With ( u ∧ v ) z = ⟨ u , z ⟩ v − ⟨ v , z ⟩ u (u\wedge v)z=\langle u,z\rangle v-\langle v,z\rangle u ( u ∧ v ) z = ⟨ u , z ⟩ v − ⟨ v , z ⟩ u , the Brown–Gray Cayley curvature formula [6 Eq. (6.12)] , in the form [26 equation (7), p. 6 of arXiv:0910.1645v1] , is
R O ( u , v ) = 3 u ∧ v + ∑ i = 0 8 S i u ∧ S i v . R_{\mathbb{O}}(u,v)=3u\wedge v+\sum_{i=0}^{8}S_i u\wedge S_i v. R O ( u , v ) = 3 u ∧ v + i = 0 ∑ 8 S i u ∧ S i v .
In our Jacobi convention this gives
J x ( R O ) = 3 ( x x T − q I ) + Z x − S h = 4 x x T − q I − 3 Q ( x ) = − Π ( x ) − 3 A ( x ) . \begin{aligned}
J_x(R_{\mathbb{O}})=3(xx^{T}-qI)+Z_x-S_h \\
&=4xx^{T}-qI-3Q(x)=-\Pi(x)-3A(x).
\end{aligned} J x ( R O ) = 3 ( x x T − q I ) + Z x − S h = 4 x x T − q I − 3 Q ( x ) = − Π ( x ) − 3 A ( x ) .
Therefore R C a y = − R O R_{\mathrm{Cay}}=-R_{\mathbb{O}} R Cay = − R O has Jacobi family Π ( x ) + 3 A ( x ) \Pi(x)+3A(x) Π ( x ) + 3 A ( x ) , with reduced eigenvalues 1 1 1 and 4 4 4 of multiplicities eight and seven.
Let C l + ( 9 ) \mathrm{Cl}_{+}(9) Cl + ( 9 ) denote the real Clifford algebra on nine generators with squares + 1 +1 + 1 and pairwise anticommutation. Then C l + ( 9 ) ≅ M 16 ( R ) ⊕ M 16 ( R ) \mathrm{Cl}_{+}(9)\cong M_{16}(\mathbb{R})\oplus M_{16}(\mathbb{R}) Cl + ( 9 ) ≅ M 16 ( R ) ⊕ M 16 ( R ) . Its representations on R 16 \mathbb{R}^{16} R 16 are equivalent up to orthogonal conjugacy and the sign of one generator; see [11 Sections 3.4–3.5, pp. 6–7] and [26 p. 6 of arXiv:0910.1645v1] . Changing a generator’s sign leaves ⟨ S i x , x ⟩ S i \langle S_i x,x\rangle S_i ⟨ S i x , x ⟩ S i unchanged. This identifies every family (1.3) with the standard Cayley family, so 3 A ( x ) = J x ( T ∗ R C a y ) − Π ( x ) 3A(x)=J_x(T^*R_{\mathrm{Cay}})-\Pi(x) 3 A ( x ) = J x ( T ∗ R Cay ) − Π ( x ) for some T ∈ O ( 16 ) T\in O(16) T ∈ O ( 16 ) and J x = ( μ − κ / 3 ) Π ( x ) + ( κ / 3 ) J x ( T ∗ R C a y ) J_x=(\mu-\kappa/3)\Pi(x)+(\kappa/3)J_x(T^*R_{\mathrm{Cay}}) J x = ( μ − κ /3 ) Π ( x ) + ( κ /3 ) J x ( T ∗ R Cay ) . As the Jacobi family determines the curvature tensor by (2.2), R = ( μ − κ / 3 ) R 1 + ( κ / 3 ) T ∗ R C a y R=(\mu-\kappa/3)R_1+(\kappa/3)T^*R_{\mathrm{Cay}} R = ( μ − κ /3 ) R 1 + ( κ /3 ) T ∗ R Cay . This finishes the proof of Theorem C: by Lemma 2.1 and § 3 , b b b satisfies (3.2) or (3.3); Propositions 5.1 and 5.2 give the forms (1.1)–(1.3), of which the first two are Clifford and the third has just been identified.
Tensors with three eigenvalues, and Jacobi-proportionality
For the standard Cayley family of § 6 , that is (1.3) with H i = S i H_i=S_i H i = S i , let E ( x ) = L ( x ) ∩ x ⊥ E(x)=L(x)\cap x^\perp E ( x ) = L ( x ) ∩ x ⊥ and N ( x ) = L ( x ) ⊥ N(x)=L(x)^\perp N ( x ) = L ( x ) ⊥ , of dimensions seven and eight; here L ( x ) = R x + Im A ( x ) L(x)=\mathbb{R}x+\operatorname{Im}A(x) L ( x ) = R x + Im A ( x ) as in § 2 .
Lemma 7.1. There is no nonzero linear map K K K with K x ∈ E ( x ) Kx\in E(x) K x ∈ E ( x ) for every x ≠ 0 x\ne0 x = 0 . There is also no nonzero linear map K K K with K x ∈ N ( x ) Kx\in N(x) K x ∈ N ( x ) for every x ≠ 0 x\ne0 x = 0 .
Proof. After the orthogonal coordinate change ( x 1 , x 2 ) ↦ ( x 1 , x ‾ 2 ) (x_1,x_2)\mapsto(x_1,\overline{x}_2) ( x 1 , x 2 ) ↦ ( x 1 , x 2 ) , the Cayley lines are
C ω = { ( u , ω u ) : u ∈ O } , ω ∈ O , C ∞ = { ( 0 , v ) : v ∈ O } . C_\omega=\{(u,\omega u):u\in\mathbb{O}\},\qquad\omega\in\mathbb{O},\qquad C_\infty=\{(0,v):v\in\mathbb{O}\}. C ω = {( u , ω u ) : u ∈ O } , ω ∈ O , C ∞ = {( 0 , v ) : v ∈ O } .
Here S ( w 0 , a ) = w 0 S 0 + ∑ i = 1 8 a i S i S_{(w_0,a)}=w_0S_0+\sum_{i=1}^{8}a_iS_i S ( w 0 , a ) = w 0 S 0 + ∑ i = 1 8 a i S i for w 0 2 + ∥ a ∥ 2 = 1 w_0^2+\lVert a\rVert^2=1 w 0 2 + ∥ a ∥ 2 = 1 and a = ∑ i a i e i a=\sum_i a_i e_i a = ∑ i a i e i . Its positive eigenspace becomes C ω C_\omega C ω with ω = a / ( 1 + w 0 ) \omega=a/(1+w_0) ω = a / ( 1 + w 0 ) when w 0 ≠ − 1 w_0\ne-1 w 0 = − 1 , and becomes C ∞ C_\infty C ∞ otherwise. Distinct lines among the C ω C_\omega C ω and C ∞ C_\infty C ∞ meet only in 0, as O \mathbb{O} O has no zero divisors, and x ∈ L ( x ) x\in L(x) x ∈ L ( x ) ; so for x x x in such a line, L ( x ) L(x) L ( x ) is that line.
Suppose first K x ∈ E ( x ) Kx\in E(x) K x ∈ E ( x ) . Since ⟨ K x , x ⟩ = 0 \langle Kx,x\rangle=0 ⟨ K x , x ⟩ = 0 , K K K is skew. Preservation of C 0 C_0 C 0 , C ∞ C_\infty C ∞ , C 1 C_1 C 1 forces K = diag ( D , D ) K=\operatorname{diag}(D,D) K = diag ( D , D ) . Preservation of C ω C_\omega C ω gives D ( ω u ) = ω D ( u ) D(\omega u)=\omega D(u) D ( ω u ) = ω D ( u ) . Put c = D ( 1 ) c=D(1) c = D ( 1 ) ; then D ( u ) = u c D(u)=uc D ( u ) = u c and ( ω u ) c = ω ( u c ) (\omega u)c=\omega(uc) ( ω u ) c = ω ( u c ) for all ω , u \omega,u ω , u . Thus c c c lies in the right nucleus of O \mathbb{O} O , which is R \mathbb{R} R . Hence K = c I K=cI K = c I , and skewness forces c = 0 c=0 c = 0 .
Suppose next K x ∈ N ( x ) Kx\in N(x) K x ∈ N ( x ) . The lines C 0 C_0 C 0 and C ∞ C_\infty C ∞ force K = ( 0 B C 0 ) K=\begin{pmatrix}0&B\\ C&0\end{pmatrix} K = ( 0 C B 0 ) . The graph condition for C ω ⊥ C_\omega^\perp C ω ⊥ is B ( ω u ) = − ω ‾ C ( u ) B(\omega u)=-\overline{\omega}C(u) B ( ω u ) = − ω C ( u ) . Taking ω = 1 \omega=1 ω = 1 gives B = − C B=-C B = − C ; taking u = 1 u=1 u = 1 then gives C ( ω ) = ω ‾ γ C(\omega)=\overline{\omega}\gamma C ( ω ) = ω γ , where γ = C ( 1 ) \gamma=C(1) γ = C ( 1 ) . Substitution yields
( a b ) γ = b ( a γ ) ( a , b ∈ O ) . (15) (ab)\gamma=b(a\gamma)\qquad(a,b\in\mathbb{O}). \tag*{(15)} ( ab ) γ = b ( aγ ) ( a , b ∈ O ) . (15)
Choose a quaternion subalgebra containing γ \gamma γ and choose noncommuting a , b a,b a , b in it. Associativity in that subalgebra gives ( a b − b a ) γ = 0 (ab-ba)\gamma=0 ( ab − ba ) γ = 0 . Multiplicativity of the octonion norm forces γ = 0 \gamma=0 γ = 0 , and therefore K = 0 K=0 K = 0 .
Proof of Theorem E . Let t ∈ { 1 , 2 } t\in\{1,2\} t ∈ { 1 , 2 } be the smallest multiplicity, with eigenvalue τ \tau τ . To apply [24 Lemma 3.1] , shift a different reduced eigenvalue to zero by subtracting a multiple of R 1 R_1 R 1 . In the convention of [24 p. 379] , the shifted tensor has k = 3 k=3 k = 3 reduced eigenvalues, one of them zero. Nikolayevsky’s [24 Lemma 3.1(1),(2), p. 383] gives an orthonormal frame for the τ \tau τ -eigenspace by odd homogeneous polynomial fields of degree at most k − 1 = 2 k-1=2 k − 1 = 2 . They are therefore linear: J 1 x , … , J t x J_1x,\ldots,J_tx J 1 x , … , J t x . Their norm and orthogonality identities make each J i J_i J i skew and orthogonal and make the J i J_i J i anticommute. Return to the original tensor, whose eigenspaces are unchanged by the shift.
Choose another reduced eigenvalue τ ′ \tau' τ ′ so that moving τ \tau τ to τ ′ \tau' τ ′ merges the multiplicities as
( 7 , 7 , 1 ) → ( 7 , 8 ) , ( 8 , 6 , 1 ) → ( 8 , 7 ) , ( 8 , 5 , 2 ) → ( 8 , 7 ) , ( 7 , 6 , 2 ) → ( 7 , 8 ) . (7,7,1)\to(7,8),\qquad(8,6,1)\to(8,7),\qquad(8,5,2)\to(8,7),\qquad(7,6,2)\to(7,8). ( 7 , 7 , 1 ) → ( 7 , 8 ) , ( 8 , 6 , 1 ) → ( 8 , 7 ) , ( 8 , 5 , 2 ) → ( 8 , 7 ) , ( 7 , 6 , 2 ) → ( 7 , 8 ) .
For a skew orthogonal J J J , let R J R_J R J denote the curvature tensor whose Jacobi family is ( J x ) ( J x ) T (Jx)(Jx)^{T} ( J x ) ( J x ) T , obtained for example from (6) . Then
R ~ = R − ( τ − τ ′ ) ∑ i = 1 t R J i \widetilde{R}=R-(\tau-\tau')\sum_{i=1}^{t}R_{J_i} R = R − ( τ − τ ′ ) i = 1 ∑ t R J i
is Osserman with two distinct eigenvalues of multiplicities seven and eight. Apply Theorem C . Its Cayley alternative is impossible: each J i x J_ix J i x remains an eigenfield in one of the two merged distributions, contradicting Lemma 7.1 after conjugating the field to the standard model: K = T J i T − 1 K=TJ_iT^{-1} K = T J i T − 1 . Thus R ~ \widetilde{R} R is Clifford.
Restore the rank-one terms successively. By [24 Lemma 3.2(1), pp. 383–384] , if R R R is Osserman, J J J is orthogonal and skew, c ≠ 0 c\ne0 c = 0 , and R − c R J R-cR_J R − c R J is Osserman and Clifford, then R R R is Clifford. Each intermediate tensor here is Osserman: only the constant eigenvalues on the mutually orthogonal eigenlines R J i x R_{J_i}x R J i x change. The lemma consequently applies at every step, in the form of Remark 7.2 whenever R − c R J R-cR_J R − c R J has a Clifford structure with ν = 8 \nu=8 ν = 8 (an endpoint of type (1.2), or the intermediate pattern ( 7 , 7 , 1 ) (7,7,1) ( 7 , 7 , 1 ) reached from ( 7 , 6 , 2 ) (7,6,2) ( 7 , 6 , 2 ) ). The sign convention of [24 ] is opposite to ours, which changes the sign of c c c ; only c ≠ 0 c\ne0 c = 0 is used. This proves the theorem.
Remark 7.2. We recall the proof of [24 Lemma 3.2(1), pp. 383–384] in the form used here. The proof in [24 ] shifts the eigenvalue of largest multiplicity of R − c R J R-cR_J R − c R J to zero; the version below shifts instead the base constant λ 0 \lambda_0 λ 0 of a Clifford structure of R − c R J R-cR_J R − c R J . The two choices differ only when ν = 8 \nu=8 ν = 8 and another eigenvalue also has multiplicity at least seven, that is, at an endpoint of type (1.2) or at the pattern ( 7 , 7 , 1 ) (7,7,1) ( 7 , 7 , 1 ) . Indeed, with the generators of zero coefficient dropped, λ 0 \lambda_0 λ 0 has multiplicity 15 − ν 15-\nu 15 − ν ; for ν ≤ 7 \nu\le7 ν ≤ 7 it is the unique eigenvalue of largest multiplicity, for ν = 8 \nu=8 ν = 8 it has multiplicity seven, and at ( 7 , 7 , 1 ) (7,7,1) ( 7 , 7 , 1 ) every Clifford structure has ν = 8 \nu=8 ν = 8 . Let J x ( R − c R J ) = λ 0 Π ( x ) + ∑ α = 1 ν μ α ( J α x ) ( J α x ) T \mathcal{J}_x(R-cR_J)=\lambda_0\Pi(x)+\sum_{\alpha=1}^{\nu}\mu_\alpha(J_\alpha x)(J_\alpha x)^T J x ( R − c R J ) = λ 0 Π ( x ) + ∑ α = 1 ν μ α ( J α x ) ( J α x ) T with μ α ≠ 0 \mu_\alpha\ne0 μ α = 0 ; subtracting λ 0 R 1 \lambda_0R_1 λ 0 R 1 from R R R we may take λ 0 = 0 \lambda_0=0 λ 0 = 0 . Let λ 1 , … , λ p \lambda_1,\ldots,\lambda_p λ 1 , … , λ p be the distinct values of the μ α \mu_\alpha μ α , and λ 0 = 0 \lambda_0=0 λ 0 = 0 . For unit x x x the eigenspaces of J x ( R − c R J ) \mathcal{J}_x(R-cR_J) J x ( R − c R J ) on x ⊥ x^\perp x ⊥ are E i ( x ) = span { J α x : μ α = λ i } E_i(x)=\operatorname{span}\{J_\alpha x:\mu_\alpha=\lambda_i\} E i ( x ) = span { J α x : μ α = λ i } for i ≥ 1 i\ge1 i ≥ 1 and E 0 ( x ) = x ⊥ ∩ ( ∑ α R J α x ) ⊥ E_0(x)=x^\perp\cap(\sum_{\alpha}R_{J_\alpha}x)^\perp E 0 ( x ) = x ⊥ ∩ ( ∑ α R J α x ) ⊥ ; let π i \pi_i π i be the orthogonal projections onto them. As J x ( R ) = J x ( R − c R J ) + c ( J x ) ( J x ) T \mathcal{J}_x(R)=\mathcal{J}_x(R-cR_J)+c(Jx)(Jx)^T J x ( R ) = J x ( R − c R J ) + c ( J x ) ( J x ) T and J x ⊥ x Jx\perp x J x ⊥ x ,
det ( t I − J x ( R ) ) det ( t I − J x ( R − c R J ) ) = 1 − c ∑ i = 0 p ∥ π i ( J x ) ∥ 2 t − λ i \frac{\det(tI-\mathcal{J}_x(R))}{\det(tI-\mathcal{J}_x(R-cR_J))}=1-c\sum_{i=0}^{p}\frac{\|\pi_i(Jx)\|^2}{t-\lambda_i} det ( t I − J x ( R − c R J )) det ( t I − J x ( R )) = 1 − c i = 0 ∑ p t − λ i ∥ π i ( J x ) ∥ 2
on x ⊥ x^\perp x ⊥ , so, both tensors being Osserman, partial fractions make each ∥ π i ( J x ) ∥ = c i \|\pi_i(Jx)\|=c_i ∥ π i ( J x ) ∥ = c i constant on the unit sphere. For i ≥ 1 i\ge1 i ≥ 1 , ∑ μ α = λ i ⟨ J α x , J x ⟩ 2 = c i 2 ∥ x ∥ 4 \sum_{\mu_\alpha=\lambda_i}\langle J_\alpha x,Jx\rangle^2=c_i^2\|x\|^4 ∑ μ α = λ i ⟨ J α x , J x ⟩ 2 = c i 2 ∥ x ∥ 4 is a sum of at most eight squares divisible by ∥ x ∥ 2 \lVert x\rVert^2 ∥ x ∥ 2 , so by [24 Lemma 3.11] each ⟨ J α x , J x ⟩ \langle J_\alpha x,Jx\rangle ⟨ J α x , J x ⟩ (a quadratic form) is a constant multiple of ∥ x ∥ 2 \lVert x\rVert^2 ∥ x ∥ 2 ; after an orthogonal change of the J α J_\alpha J α with μ α = λ i \mu_\alpha=\lambda_i μ α = λ i we get π i ( J x ) = c i J ( i ) x \pi_i(Jx)=c_iJ^{(i)}x π i ( J x ) = c i J ( i ) x for one of them, J ( i ) J^{(i)} J ( i ) . Then J ν + 1 x : = π 0 ( J x ) = J x − ∑ i ≥ 1 c i J ( i ) x J_{\nu+1}x:=\pi_0(Jx)=Jx-\sum_{i\geq1}c_iJ^{(i)}x J ν + 1 x := π 0 ( J x ) = J x − ∑ i ≥ 1 c i J ( i ) x is linear and skew, anticommutes with every J α J_\alpha J α (as J ν + 1 x ⊥ J α x J_{\nu+1}x\perp J_\alpha x J ν + 1 x ⊥ J α x ), and ∥ J ν + 1 x ∥ 2 = c 0 2 ∥ x ∥ 2 \lVert J_{\nu+1}x\rVert^2=c_0^2\lVert x\rVert^2 ∥ J ν + 1 x ∥ 2 = c 0 2 ∥ x ∥ 2 with c 0 2 = 1 − ∑ i > 1 c i 2 c_0^2=1-\sum_{i>1}c_i^2 c 0 2 = 1 − ∑ i > 1 c i 2 . If ν = 8 \nu=8 ν = 8 , then c 0 = 0 c_0=0 c 0 = 0 : otherwise c 0 − 1 J 9 c_0^{-1}J_9 c 0 − 1 J 9 would be a ninth orthogonal complex structure on R 16 \mathbb{R}^{16} R 16 anticommuting with J 1 , … , J 8 J_1,\ldots,J_8 J 1 , … , J 8 , and R 16 \mathbb{R}^{16} R 16 carries at most eight (Hurwitz–Radon). Finally, write J x = ∑ i ≥ 1 c i J ( i ) x + c 0 ( c 0 − 1 J ν + 1 x ) Jx=\sum_{i\geq1}c_iJ^{(i)}x+c_0(c_0^{-1}J_{\nu+1}x) J x = ∑ i ≥ 1 c i J ( i ) x + c 0 ( c 0 − 1 J ν + 1 x ) , omitting the last term if c 0 = 0 c_0=0 c 0 = 0 . An orthogonal change of J ( 1 ) , … , J ( p ) J^{(1)},\ldots,J^{(p)} J ( 1 ) , … , J ( p ) (and of c 0 − 1 J ν + 1 c_0^{-1}J_{\nu+1} c 0 − 1 J ν + 1 if c 0 ≠ 0 c_0\neq0 c 0 = 0 ) that diagonalizes diag ( λ 1 , … , λ p , 0 ) + c γ γ T \operatorname{diag}(\lambda_1,\ldots,\lambda_p,0)+c\gamma\gamma^{T} diag ( λ 1 , … , λ p , 0 ) + c γ γ T , γ = ( c 1 , … , c p , c 0 ) T \gamma=(c_1,\ldots,c_p,c_0)^{T} γ = ( c 1 , … , c p , c 0 ) T (without the last row and column if c 0 = 0 c_0=0 c 0 = 0 ), gives a Clifford structure for R R R .
Proof of Corollary D. Andrejić [1 Theorem 5] , restated as [3 Theorem 3, p. 8470] , proves that a Riemannian Jacobi-proportional Osserman algebraic curvature tensor on R 16 \mathbb{R}^{16} R 16 that is not Clifford has multiplicity pattern ( 8 , 7 ) (8,7) ( 8 , 7 ) or ( 7 , 7 , 1 ) (7,7,1) ( 7 , 7 , 1 ) . Andrejić’s hypothesis places no restriction on the number of eigenvalues. Apply Theorems C and E . □
Further multiplicity reductions
Theorems C and E complete several polynomial-eigenfield reductions of [24 ] .
Theorem F. An Osserman algebraic curvature tensor on R 16 \mathbb{R}^{16} R 16 is Clifford if its reduced multiplicity pattern is one of
( 7 , 6 , 1 , 1 ) , ( 8 , 5 , 1 , 1 ) , ( 7 , 5 , 2 , 1 ) , ( 8 , 4 , 2 , 1 ) , ( 7 , 5 , 1 , 1 , 1 ) , ( 8 , 4 , 1 , 1 , 1 ) , (7,6,1,1),\ (8,5,1,1),\ (7,5,2,1),\ (8,4,2,1),\ (7,5,1,1,1),\ (8,4,1,1,1), ( 7 , 6 , 1 , 1 ) , ( 8 , 5 , 1 , 1 ) , ( 7 , 5 , 2 , 1 ) , ( 8 , 4 , 2 , 1 ) , ( 7 , 5 , 1 , 1 , 1 ) , ( 8 , 4 , 1 , 1 , 1 ) ,
or if one eigenvalue has multiplicity seven or eight and every other eigenvalue has multiplicity at most two (nine further patterns, fifteen in all).
Constant projections onto Cayley eigenspaces
Lemma 8.1. Let E ( x ) , N ( x ) E(x),N(x) E ( x ) , N ( x ) be the two Cayley distributions of Lemma 7.1 . If a skew-symmetric linear map K K K satisfies ∥ π N ( x ) K x ∥ 2 = c ∥ x ∥ 2 \lVert\pi_{N(x)}Kx\rVert^2=c\lVert x\rVert^2 ∥ π N ( x ) K x ∥ 2 = c ∥ x ∥ 2 for all x ≠ 0 x\neq0 x = 0 and a constant c c c , then K = 0 K=0 K = 0 . In particular an orthogonal skew map cannot have constant normalized projection lengths onto the two Cayley eigenspaces.
Proof. Use the symmetric Clifford system of § 6 , written here H 0 , … , H 8 H_0,\ldots,H_8 H 0 , … , H 8 (there S 0 , … , S 8 S_0,\ldots,S_8 S 0 , … , S 8 ), and put q = ∥ x ∥ 2 q=\lVert x\rVert^2 q = ∥ x ∥ 2 , h i ( x ) = ⟨ H i x , x ⟩ h_i(x)=\langle H_i x,x\rangle h i ( x ) = ⟨ H i x , x ⟩ , a i ( x ) = ⟨ H i x , K x ⟩ a_i(x)=\langle H_i x,Kx\rangle a i ( x ) = ⟨ H i x , K x ⟩ . The Clifford relations give ⟨ H i x , H j x ⟩ = δ i j q \langle H_i x,H_jx\rangle=\delta_{ij}q ⟨ H i x , H j x ⟩ = δ ij q . The identity ∑ i ( H i x ) ( H i x ) T = Z x = x x T + q I − Q ( x ) \sum_i(H_i x)(H_i x)^{T}=Z_x=xx^{T}+qI-Q(x) ∑ i ( H i x ) ( H i x ) T = Z x = x x T + q I − Q ( x ) of § 6 (with H i = S i H_i=S_i H i = S i ) and Q ( x ) = q π L ( x ) Q(x)=q\pi_{L(x)} Q ( x ) = q π L ( x ) (§ 2 ), with x ∈ L ( x ) x\in L(x) x ∈ L ( x ) , give ∑ i ( H i x ) ( H i x ) T = x x T + q π N ( x ) = q π R x + N ( x ) \sum_i(H_i x)(H_i x)^{T}=xx^{T}+q\pi_{N(x)}=q\pi_{\mathbb{R}x+N(x)} ∑ i ( H i x ) ( H i x ) T = x x T + q π N ( x ) = q π R x + N ( x ) . So the nine orthogonal vectors H i x H_i x H i x , of norm q \sqrt q q , span R x + N ( x ) \mathbb{R}x+N(x) R x + N ( x ) , and, K K K being skew, ∑ i a i 2 = c q 2 \sum_i a_i^2=cq^2 ∑ i a i 2 = c q 2 and ∑ i h i 2 = q 2 \sum_i h_i^2=q^2 ∑ i h i 2 = q 2 ; differentiating the latter along K x Kx K x gives ∑ i h i a i = 0 \sum_i h_i a_i=0 ∑ i h i a i = 0 . The matrix of a i a_i a i is the symmetric trace-free A i = ( H i K − K H i ) / 2 A_i=(H_iK-KH_i)/2 A i = ( H i K − K H i ) /2 .
If c > 0 c>0 c > 0 , put H i ( t ) = c H i + c − 1 / 2 sin t A i H_i(t)=cH_i+c^{-1/2}\sin t A_i H i ( t ) = c H i + c − 1/2 sin t A i . Then ∑ i ⟨ H i ( t ) x , x ⟩ 2 = q 2 \sum_i\langle H_i(t)x,x\rangle^2=q^2 ∑ i ⟨ H i ( t ) x , x ⟩ 2 = q 2 for every t t t , and this quadratic map is not constant on the unit sphere (otherwise H i ( t ) = b i I H_i(t)=b_iI H i ( t ) = b i I with b b b a unit vector, but all its matrices are trace-free). By Lemma 4.1(c) the H i ( t ) H_i(t) H i ( t ) form a symmetric Clifford system, so Ω ( t ) = H 0 ( t ) ⋯ H 8 ( t ) \Omega(t)=H_0(t)\cdots H_8(t) Ω ( t ) = H 0 ( t ) ⋯ H 8 ( t ) is a symmetric involution, and tr Ω ( t ) \operatorname{tr}\Omega(t) tr Ω ( t ) is continuous and integer-valued, hence constant. But Ω ( 0 ) = ± I \Omega(0)=\pm I Ω ( 0 ) = ± I on R 16 \mathbb{R}^{16} R 16 (as nine is odd, Ω ( 0 ) \Omega(0) Ω ( 0 ) commutes with every H i H_i H i , and, by § 6 , the products of the H i H_i H i span the symmetric matrices, so a matrix commuting with every H i H_i H i is scalar) and Ω ( π ) = − Ω ( 0 ) \Omega(\pi)=-\Omega(0) Ω ( π ) = − Ω ( 0 ) , a contradiction. If c = 0 c=0 c = 0 , then K x ∈ L ( x ) Kx\in L(x) K x ∈ L ( x ) , and skewness gives K x ∈ E ( x ) Kx\in E(x) K x ∈ E ( x ) , so K = 0 K=0 K = 0 by Lemma 7.1 . □
Polynomial tools and a three-column reduction
From here to the end of this section, A ( x ) A(x) A ( x ) denotes the Jacobi operator J x \mathcal{J}_x J x after one reduced eigenvalue has been shifted to zero (not the normalized family of Section 2), q = ∥ x ∥ 2 q=\lVert x\rVert^2 q = ∥ x ∥ 2 , O = R [ x 1 , … , x 16 ] / ( q ) \mathcal{O}=\mathbb{R}[x_1,\ldots,x_{16}]/(q) O = R [ x 1 , … , x 16 ] / ( q ) , and π : R [ x 1 , … , x 16 ] → O \pi:\mathbb{R}[x_1,\ldots,x_{16}]\to\mathcal{O} π : R [ x 1 , … , x 16 ] → O is the projection. As q q q is irreducible, it is a prime element of R [ x 1 , … , x 16 ] \mathbb{R}[x_1,\ldots,x_{16}] R [ x 1 , … , x 16 ] . We use two results of [24 ] : O \mathcal{O} O is a UFD [24 Theorem 3.1, p. 397] ; and if q q q divides a sum of at most eight real polynomial squares, it divides every polynomial being squared [24 Lemma 3.11, p. 398] .
A polynomial eigenvector field of a simple eigenvalue is a simple eigenfield . The simple eigenfields of [24 Lemma 3.1(1), p. 383] satisfy
⟨ P ( x ) , x ⟩ = 0 , ∥ P ( x ) ∥ 2 = q 3 , P ( P ( x ) ) = − q 4 x , (16) \langle P(x),x\rangle=0,\qquad\lVert P(x)\rVert^2=q^3,\qquad P(P(x))=-q^4x, \tag*{(16)} ⟨ P ( x ) , x ⟩ = 0 , ∥ P ( x ) ∥ 2 = q 3 , P ( P ( x )) = − q 4 x , (16)
after multiplication by q q q if they are linear: equation (7) of that lemma states them for unit x x x , and homogeneity gives them for all x x x . We also use [24 Lemma 3.5, pp. 388–390] . It is stated there under the standing hypothesis of [24 Subsection 3.1] , a Jacobi eigenvalue of multiplicity nine, and for an eigenvector field P P P ; its proof uses neither. We state it in the form used here, with that proof.
Lemma 8.2. Let P : R 16 → R 16 P:\mathbb{R}^{16}\to\mathbb{R}^{16} P : R 16 → R 16 have homogeneous cubic components that satisfy (8.1). If the images in O \mathcal{O} O of the sixteen components of P P P have a nonunit common divisor, then P ( x ) = q J x P(x)=qJx P ( x ) = q J x for an orthogonal skew map J J J .
Proof. We follow [24 pp. 389–390] . If every component of P P P is divisible by q q q , then P ( x ) = q J x P(x)=qJx P ( x ) = q J x with J J J linear, and (8.1) gives ⟨ J x , x ⟩ = 0 \langle Jx,x\rangle=0 ⟨ J x , x ⟩ = 0 and ∥ J x ∥ 2 = q \lVert Jx\rVert^2=q ∥ J x ∥ 2 = q , so J J J is skew and orthogonal. Otherwise the common divisor is homogeneous, as a divisor of a nonzero homogeneous element of the graded domain O \mathcal{O} O ; its degree r r r is positive, since the units of O \mathcal{O} O are the nonzero constants, and r ≤ 3 r\leq3 r ≤ 3 , since it divides a nonzero cubic. Lifting to R [ x 1 , … , x 16 ] \mathbb{R}[x_1,\ldots,x_{16}] R [ x 1 , … , x 16 ] , we get P ( x ) = φ ( x ) Ψ ( x ) + q C x P(x)=\varphi(x)\Psi(x)+qCx P ( x ) = φ ( x ) Ψ ( x ) + q C x with φ \varphi φ homogeneous of degree r r r and not divisible by q q q , the components of Ψ \Psi Ψ homogeneous of degree 3 − r 3-r 3 − r , and C C C linear. It suffices to show that every component of Ψ \Psi Ψ is divisible by q q q .
If r = 3 r=3 r = 3 , then Ψ \Psi Ψ is constant and φ ⟨ Ψ , x ⟩ = − q ⟨ C x , x ⟩ \varphi\langle\Psi,x\rangle=-q\langle Cx,x\rangle φ ⟨ Ψ , x ⟩ = − q ⟨ C x , x ⟩ . As q q q is prime and does not divide φ \varphi φ , it divides the linear form ⟨ Ψ , x ⟩ \langle\Psi,x\rangle ⟨ Ψ , x ⟩ , so Ψ = 0 \Psi=0 Ψ = 0 .
If r = 2 r=2 r = 2 , then Ψ ( x ) = B x \Psi(x)=Bx Ψ ( x ) = B x with B B B linear. Modulo q q q , ∥ P ( x ) ∥ 2 ≡ φ ( x ) 2 ∥ B x ∥ 2 \lVert P(x)\rVert^2\equiv\varphi(x)^2\lVert Bx\rVert^2 ∥ P ( x ) ∥ 2 ≡ φ ( x ) 2 ∥ B x ∥ 2 , so q q q divides ∥ B x ∥ 2 \lVert Bx\rVert^2 ∥ B x ∥ 2 , and B = c U B=cU B = c U with U U U orthogonal and c ≥ 0 c\geq0 c ≥ 0 . Since ∥ P ( x ) ∥ 2 = q 3 \lVert P(x)\rVert^2=q^3 ∥ P ( x ) ∥ 2 = q 3 and P ( x ) ≡ φ ( x ) B x P(x)\equiv\varphi(x)Bx P ( x ) ≡ φ ( x ) B x modulo q q q , we get
0 ≡ P ( P ( x ) ) ≡ φ ( P ( x ) ) B P ( x ) ≡ φ ( x ) 3 φ ( B x ) B 2 x ( m o d q ) . 0\equiv P(P(x))\equiv\varphi(P(x))BP(x)\equiv\varphi(x)^3\varphi(Bx)B^2x\pmod q. 0 ≡ P ( P ( x )) ≡ φ ( P ( x )) B P ( x ) ≡ φ ( x ) 3 φ ( B x ) B 2 x ( mod q ) .
If c ≠ 0 c\neq0 c = 0 , then B 2 B^2 B 2 is invertible, so q q q divides φ ( B x ) \varphi(Bx) φ ( B x ) , and the substitution x = B − 1 y x=B^{-1}y x = B − 1 y shows that c − 2 ∥ y ∥ 2 c^{-2}\lVert y\rVert^2 c − 2 ∥ y ∥ 2 divides φ ( y ) \varphi(y) φ ( y ) , which is false. So c = 0 c=0 c = 0 and Ψ = 0 \Psi=0 Ψ = 0 .
If r = 1 r=1 r = 1 , then φ \varphi φ is a nonzero linear form, and φ ⟨ Ψ , x ⟩ = − q ⟨ C x , x ⟩ \varphi\langle\Psi,x\rangle=-q\langle Cx,x\rangle φ ⟨ Ψ , x ⟩ = − q ⟨ C x , x ⟩ shows that φ \varphi φ divides ⟨ C x , x ⟩ \langle Cx,x\rangle ⟨ C x , x ⟩ , say ⟨ C x , x ⟩ = φ ( x ) ⟨ a , x ⟩ \langle Cx,x\rangle=\varphi(x)\langle a,x\rangle ⟨ C x , x ⟩ = φ ( x ) ⟨ a , x ⟩ with a ∈ R 16 a\in\mathbb{R}^{16} a ∈ R 16 . Replacing C C C by C ~ = C − φ ( x ) a \widetilde{C}=C-\varphi(x)a C = C − φ ( x ) a and Ψ \Psi Ψ by Ψ ~ = Ψ + q a \widetilde{\Psi}=\Psi+qa Ψ = Ψ + q a keeps P = φ Ψ ~ + q C ~ x P=\varphi\widetilde{\Psi}+q\widetilde{C}x P = φ Ψ + q C x and makes C ~ \widetilde{C} C skew. Modulo φ \varphi φ , the identity ∥ P ( x ) ∥ 2 = q 3 \lVert P(x)\rVert^2=q^3 ∥ P ( x ) ∥ 2 = q 3 shows that φ \varphi φ divides q − ∥ C ~ x ∥ 2 = x T ( I + C ~ 2 ) x q-\lVert\widetilde{C}x\rVert^2=x^{T}(I+\widetilde{C}^{2})x q − ∥ C x ∥ 2 = x T ( I + C 2 ) x . A quadratic form divisible by a linear form has rank at most two and vanishes at a nonzero vector of every plane, so it is definite on no plane; as every eigenvalue of the symmetric matrix I + C ~ 2 I+\widetilde{C}^{2} I + C 2 has even multiplicity, I + C ~ 2 = 0 I+\widetilde{C}^{2}=0 I + C 2 = 0 , and C ~ \widetilde{C} C is orthogonal. The third identity in (8.1) now reads φ ( P ( x ) ) Ψ ~ ( P ( x ) ) = − q 3 φ ( x ) C ~ Ψ ~ ( x ) \varphi(P(x))\widetilde{\Psi}(P(x))=-q^3\varphi(x)\widetilde{C}\widetilde{\Psi}(x) φ ( P ( x )) Ψ ( P ( x )) = − q 3 φ ( x ) C Ψ ( x ) . On the hyperplane φ = 0 \varphi=0 φ = 0 we have P ( x ) = q C ~ x P(x)=q\widetilde{C}x P ( x ) = q C x , so φ ( C ~ x ) Ψ ~ ( C ~ x ) \varphi(\widetilde{C}x)\widetilde{\Psi}(\widetilde{C}x) φ ( C x ) Ψ ( C x ) vanishes there. Writing φ ( x ) = ⟨ f , x ⟩ \varphi(x)=\langle f,x\rangle φ ( x ) = ⟨ f , x ⟩ , the form φ ( C ~ x ) = − ⟨ C ~ f , x ⟩ \varphi(\widetilde{C}x)=-\langle\widetilde{C}f,x\rangle φ ( C x ) = − ⟨ C f , x ⟩ does not vanish identically on f ⊥ f^\perp f ⊥ , since C ~ f ≠ 0 \widetilde{C}f\neq0 C f = 0 is orthogonal to f f f . Hence φ \varphi φ divides every component of Ψ ~ ( C ~ x ) \widetilde{\Psi}(\widetilde{C}x) Ψ ( C x ) , and, as C ~ − 1 = − C ~ \widetilde{C}^{-1}=-\widetilde{C} C − 1 = − C , we get Ψ ~ ( x ) = φ ( C ~ x ) B ′ x \widetilde{\Psi}(x)=\varphi(\widetilde{C}x)B'x Ψ ( x ) = φ ( C x ) B ′ x with B ′ B' B ′ linear. Then P ( x ) = φ ( x ) φ ( C ~ x ) B ′ x + q C ~ x P(x)=\varphi(x)\varphi(\widetilde{C}x)B'x+q\widetilde{C}x P ( x ) = φ ( x ) φ ( C x ) B ′ x + q C x , where φ ( x ) φ ( C ~ x ) \varphi(x)\varphi(\widetilde{C}x) φ ( x ) φ ( C x ) is a quadratic form not divisible by q q q , and the case r = 2 r=2 r = 2 applies: B ′ = 0 B'=0 B ′ = 0 , so Ψ ~ = 0 \widetilde{\Psi}=0 Ψ = 0 , and Ψ = − q a \Psi=-qa Ψ = − q a is divisible by q q q . □
The three-column step of [24 pp. 387–389] reads as follows. Let P 1 , P 2 , P 3 P_1,P_2,P_3 P 1 , P 2 , P 3 be mutually orthogonal cubic eigenfields of squared norm q 3 q^3 q 3 satisfying ∑ i = 1 3 w i P i P i T ≡ 0 \sum_{i=1}^3 w_iP_iP_i^{T}\equiv0 ∑ i = 1 3 w i P i P i T ≡ 0 modulo q q q , with w 1 , w 2 > 0 > w 3 w_1,w_2>0>w_3 w 1 , w 2 > 0 > w 3 (identities of this kind come from the spectral projectors, as in (19) below). After rescaling the columns by nonzero constants, their images p i ∈ O 16 p_i\in\mathcal{O}^{16} p i ∈ O 16 satisfy
p 1 p 1 T + p 2 p 2 T = p 3 p 3 T . (17) p_1p_1^{T}+p_2p_2^{T}=p_3p_3^{T}. \tag*{(17)} p 1 p 1 T + p 2 p 2 T = p 3 p 3 T . (17)
The right side of (17) has rank at most one, so every 2 × 2 2\times2 2 × 2 minor of the left side vanishes; by Lagrange’s identity its principal minor in rows and columns a , b a,b a , b is ( p 1 a p 2 b − p 2 a p 1 b ) 2 (p_{1a}p_{2b}-p_{2a}p_{1b})^2 ( p 1 a p 2 b − p 2 a p 1 b ) 2 , so, O \mathcal{O} O being a domain, all the minors p 1 a p 2 b − p 2 a p 1 b p_{1a}p_{2b}-p_{2a}p_{1b} p 1 a p 2 b − p 2 a p 1 b of [ p 1 p 2 ] [p_1\ p_2] [ p 1 p 2 ] vanish. If some p i = 0 p_i=0 p i = 0 , then P i P_i P i is q q q times a linear eigenfield. Otherwise unique factorization gives a primitive v v v with p 1 = u 1 v , p 2 = u 2 v p_1=u_1v,p_2=u_2v p 1 = u 1 v , p 2 = u 2 v ; then (17) gives p 3 = u 3 v p_3=u_3v p 3 = u 3 v and u 1 2 + u 2 2 = u 3 2 u_1^2+u_2^2=u_3^2 u 1 2 + u 2 2 = u 3 2 : a nonzero component of v v v gives a square root of u 1 2 + u 2 2 u_1^2+u_2^2 u 1 2 + u 2 2 in the fraction field, integral over the integrally closed ring O \mathcal{O} O , hence in O \mathcal{O} O ; cross entries fix one sign in all components. Suppose the negative column comes from a simple eigenvalue. If u 3 u_3 u 3 is a nonunit, Lemma 8.2 makes P 3 / q P_3/q P 3 / q linear. If u 3 u_3 u 3 is a unit, homogeneity makes p 1 , p 2 p_1,p_2 p 1 , p 2 constant multiples of p 3 p_3 p 3 , and a constant combination of the original fields satisfies
c 1 P 1 ( x ) + c 2 P 2 ( x ) = q J x , c 1 2 + c 2 2 = 1 , (18) c_1P_1(x)+c_2P_2(x)=qJx,\qquad c_1^2+c_2^2=1, \tag*{(18)} c 1 P 1 ( x ) + c 2 P 2 ( x ) = q J x , c 1 2 + c 2 2 = 1 , (18)
with J J J orthogonal and skew (by norm and tangency). If P 1 , P 2 P_1,P_2 P 1 , P 2 have the same eigenvalue, this is a linear eigenfield. Otherwise J x Jx J x has constant coordinates in their moving orthonormal frame on the unit sphere, so subtracting C R J CR_J C R J , for any constant C C C , leaves an Osserman tensor.
Two simple eigenvalues
Consider ( M , N , 1 , 1 ) (M,N,1,1) ( M , N , 1 , 1 ) with ( M , N ) = ( 7 , 6 ) (M,N)=(7,6) ( M , N ) = ( 7 , 6 ) or ( 8 , 5 ) (8,5) ( 8 , 5 ) . Shift the eigenvalue of multiplicity M M M to zero; let λ \lambda λ be the eigenvalue of multiplicity N N N and μ , σ \mu,\sigma μ , σ the simple eigenvalues (distinct and nonzero). By [24 Lemma 3.1] there are cubic normalized eigenfields P 1 , P 2 P_1,P_2 P 1 , P 2 (linear ones homogenized by q q q ), and the spectral projectors give
μ ( μ − λ ) P 1 P 1 T + σ ( σ − λ ) P 2 P 2 T = q A ( A − λ q I ) . (19) \mu(\mu-\lambda)P_1P_1^{T}+\sigma(\sigma-\lambda)P_2P_2^{T}=qA(A-\lambda qI). \tag*{(19)} μ ( μ − λ ) P 1 P 1 T + σ ( σ − λ ) P 2 P 2 T = q A ( A − λ q I ) . (19)
If the two coefficients have the same sign, the sum-of-squares lemma on the diagonal makes both fields q q q times linear ones. Moving one simple eigenvalue into each large eigenvalue gives the pattern ( M + 1 , N + 1 ) (M+1,N+1) ( M + 1 , N + 1 ) , that is ( 8 , 7 ) (8,7) ( 8 , 7 ) or ( 9 , 6 ) (9,6) ( 9 , 6 ) : the first is classified by Theorem C, with Cayley excluded by Lemma 7.1; the second is Clifford by [24 Proposition 2.1, p. 382] . Restore the two rank-one terms with [24 Lemma 3.2(1)] , first the one moved into λ \lambda λ and then the one moved into 0 0 0 . A Clifford structure with ν = 8 \nu=8 ν = 8 (the J i J_i J i of η i = 0 \eta_i=0 η i = 0 dropped) gives its base eigenvalue multiplicity exactly 15 − ν = 7 15-\nu=7 15 − ν = 7 , the dimension of x ⊥ ∩ ( ∑ i R J i x ) ⊥ x^\perp\cap(\sum_iR_{J_i}x)^\perp x ⊥ ∩ ( ∑ i R J i x ) ⊥ . The intermediate pattern ( M + 1 , N , 1 ) (M+1,N,1) ( M + 1 , N , 1 ) , that is ( 8 , 6 , 1 ) (8,6,1) ( 8 , 6 , 1 ) or ( 9 , 5 , 1 ) (9,5,1) ( 9 , 5 , 1 ) , has no eigenvalue of multiplicity seven, so its Clifford structures have ν ≤ 7 \nu\leq7 ν ≤ 7 , and Remark 7.2 is needed only at an endpoint of type ( 1.2 ) (1.2) ( 1.2 ) . This proves the claim.
Otherwise, after interchanging μ , σ \mu,\sigma μ , σ , put a 2 = μ ( μ − λ ) > 0 a^2=\mu(\mu-\lambda)>0 a 2 = μ ( μ − λ ) > 0 and b 2 = − σ ( σ − λ ) > 0 b^2=-\sigma(\sigma-\lambda)>0 b 2 = − σ ( σ − λ ) > 0 , and let p i p_i p i be the image of the unscaled P i P_i P i in O 16 \mathcal{O}^{16} O 16 . Modulo q q q , (19) gives ( a p 1 ) ( a p 1 ) T = ( b p 2 ) ( b p 2 ) T (ap_1)(ap_1)^T=(bp_2)(bp_2)^T ( a p 1 ) ( a p 1 ) T = ( b p 2 ) ( b p 2 ) T , so a p 1 = ± b p 2 ap_1=\pm bp_2 a p 1 = ± b p 2 (compare a nonzero component, then every cross entry). If both vanish, the linear-field argument above applies. Otherwise, after changing the sign of P 2 P_2 P 2 ,
a P 1 − b P 2 = a 2 + b 2 q J x aP_1-bP_2=\sqrt{a^2+b^2}\,qJx a P 1 − b P 2 = a 2 + b 2 q J x
with J J J orthogonal and skew. Put C = μ + σ − λ C=\mu+\sigma-\lambda C = μ + σ − λ ; it is nonzero since a 2 + b 2 = ( μ − σ ) C a^2+b^2=(\mu-\sigma)C a 2 + b 2 = ( μ − σ ) C . On the unit sphere, subtracting C ( J x ) ( J x ) T C(Jx)(Jx)^T C ( J x ) ( J x ) T turns the two-dimensional block into a fixed symmetric matrix of trace λ \lambda λ and determinant μ σ − C ( σ a 2 + μ b 2 ) / ( a 2 + b 2 ) = 0 \mu\sigma-C(\sigma a^2+\mu b^2)/(a^2+b^2)=0 μ σ − C ( σ a 2 + μ b 2 ) / ( a 2 + b 2 ) = 0 , since σ a 2 + μ b 2 = μ σ ( μ − σ ) \sigma a^2+\mu b^2=\mu\sigma(\mu-\sigma) σ a 2 + μ b 2 = μ σ ( μ − σ ) . The new eigenvalues are 0 , λ 0,\lambda 0 , λ , so the pattern is ( M + 1 , N + 1 ) (M+1,N+1) ( M + 1 , N + 1 ) , and the projection lengths of J x Jx J x onto the new eigenspaces are constant, because the changed block is a constant matrix in the frame ( P 1 , P 2 ) (P_1,P_2) ( P 1 , P 2 ) . Lemma 8.1 excludes a Cayley endpoint (for ( M , N ) = ( 8 , 5 ) (M,N)=(8,5) ( M , N ) = ( 8 , 5 ) the endpoint ( 9 , 6 ) (9,6) ( 9 , 6 ) is Clifford by [24 Proposition 2.1] ). So the endpoint is Clifford, and [24 Lemma 3.2(1)] , with Remark 7.2 at an endpoint of type (1.2), restores the original tensor.
A double and a simple eigenvalue
Consider ( M , N , 2 , 1 ) (M,N,2,1) ( M , N , 2 , 1 ) with ( M , N ) = ( 7 , 5 ) (M,N)=(7,5) ( M , N ) = ( 7 , 5 ) or ( 8 , 4 ) (8,4) ( 8 , 4 ) . Shift M M M to zero and call the other eigenvalues λ , μ , σ \lambda,\mu,\sigma λ , μ , σ , with μ \mu μ double and σ \sigma σ simple. There are cubic normalized fields P 1 , P 2 P_1,P_2 P 1 , P 2 for μ \mu μ and P 3 P_3 P 3 for σ \sigma σ with
μ ( μ − λ ) ( P 1 P 1 T + P 2 P 2 T ) + σ ( σ − λ ) P 3 P 3 T = q A ( A − λ q I ) . \mu(\mu-\lambda)(P_1P_1^{T}+P_2P_2^{T})+\sigma(\sigma-\lambda)P_3P_3^{T}=qA(A-\lambda qI). μ ( μ − λ ) ( P 1 P 1 T + P 2 P 2 T ) + σ ( σ − λ ) P 3 P 3 T = q A ( A − λ q I ) .
For equal coefficient signs all fields are linear after division by q q q . For opposite signs, after multiplying by − 1 -1 − 1 if necessary, the two double-eigenvalue columns have positive coefficients and the simple column the negative one, and the three-column reduction applies: its nonunit case gives a linear simple-eigenvalue field, its unit case a linear double-eigenvalue field (8.3), and a column vanishing modulo q q q is already a linear field. Moving that field into the zero eigenvalue gives the endpoints
Table 2.
which are Clifford by Theorem E , Subsection 8.3 or [24 Proposition 2.1, p. 382] . Restore the rank-one term by [24 Lemma 3.2(1)] ; no endpoint in the table has an eigenvalue of multiplicity seven, so their Clifford structures have ν ≤ 7 \nu\leq7 ν ≤ 7 and Remark 7.2 is not needed.
Three simple eigenvalues
Consider ( M , N , 1 , 1 , 1 ) (M,N,1,1,1) ( M , N , 1 , 1 , 1 ) with ( M , N ) = ( 7 , 5 ) (M,N)=(7,5) ( M , N ) = ( 7 , 5 ) or ( 8 , 4 ) (8,4) ( 8 , 4 ) . After shifting M M M to zero there are five reduced eigenvalues, so [24 Lemma 3.1] gives fields of odd degree at most three. Let λ \lambda λ be the eigenvalue of multiplicity N N N and μ 1 , μ 2 , μ 3 \mu_1,\mu_2,\mu_3 μ 1 , μ 2 , μ 3 the simple eigenvalues; the cubic spectral identity has weights w i = μ i ( μ i − λ ) w_i=\mu_i(\mu_i-\lambda) w i = μ i ( μ i − λ ) . If all signs agree, every field is linear after division by q q q , and any one of them serves as the linear eigenfield below. Otherwise choose the sign and order so that two weights are positive and the third, a simple column, is negative; the three-column reduction applies.
If we have a linear eigenfield, moving it into the zero eigenvalue gives ( 8 , 5 , 1 , 1 ) (8,5,1,1) ( 8 , 5 , 1 , 1 ) or ( 9 , 4 , 1 , 1 ) (9,4,1,1) ( 9 , 4 , 1 , 1 ) , Clifford by Subsection 8.3 and [24 Proposition 2.1] ; restore by [24 Lemma 3.2(1)] . Otherwise (8.3) mixes μ 1 , μ 2 \mu_1,\mu_2 μ 1 , μ 2 . Put d = μ 1 2 + μ 2 2 d=\mu_1^2+\mu_2^2 d = μ 1 2 + μ 2 2 . After subtracting C ( J x ) ( J x ) T C(Jx)(Jx)^T C ( J x ) ( J x ) T the characteristic equation of the block is ( μ 1 − z ) ( μ 2 − z ) − C ( d − z ) = 0 (\mu_1-z)(\mu_2-z)-C(d-z)=0 ( μ 1 − z ) ( μ 2 − z ) − C ( d − z ) = 0 . Choose z ∈ { 0 , λ } z\in\{0,\lambda\} z ∈ { 0 , λ } with d ≠ z d\ne z d = z (possible since λ ≠ 0 \lambda\ne0 λ = 0 ) and C = ( μ 1 − z ) ( μ 2 − z ) / ( d − z ) C=(\mu_1-z)(\mu_2-z)/(d-z) C = ( μ 1 − z ) ( μ 2 − z ) / ( d − z ) , which is nonzero. The other eigenvalue is θ = μ 1 + μ 2 − C − z \theta=\mu_1+\mu_2-C-z θ = μ 1 + μ 2 − C − z , and 0 , λ , μ 3 0,\lambda,\mu_3 0 , λ , μ 3 survive unchanged and distinct, so there is no two-eigenvalue endpoint. The complete list of endpoints, including coincidences, is as follows, where other large means θ = λ \theta=\lambda θ = λ if z = 0 z=0 z = 0 and θ = 0 \theta=0 θ = 0 if z = λ z=\lambda z = λ :
Table 3.
Each is Clifford by the preceding subsections, Theorem E, [24 Proposition 2.1] (an eigenvalue of multiplicity 9), or [23 Proposition 1] for ( 10 , 4 , 1 ) (10,4,1) ( 10 , 4 , 1 ) (in the notation of [23 Proposition 1] , m = 10 m=10 m = 10 and ν = 5 \nu=5 ν = 5 satisfy 16 ≥ 3 ν 16 \ge3\nu 16 ≥ 3 ν and 16 > ( ν + 1 ) 2 / 4 16 > (\nu+1)^2/4 16 > ( ν + 1 ) 2 /4 ). Each update has constant eigenvalues, so [24 Lemma 3.2(1)] , in the form of Remark 7.2 at the endpoint ( 7 , 7 , 1 ) (7,7,1) ( 7 , 7 , 1 ) , restores the original tensor.
The remaining patterns with multiplicities at most two
Shift the eigenvalue of multiplicity M ∈ { 7 , 8 } M \in\{7,8\} M ∈ { 7 , 8 } to zero and let r = 15 − M ≤ 8 r=15-M \le8 r = 15 − M ≤ 8 . By [24 Lemma 3.1, p. 383] , which applies because every nonzero eigenvalue has multiplicity one or two, there are r r r orthogonal normalized fields for the nonzero eigenvalues. Divide each by the largest possible power of q q q and homogenize all to the largest remaining odd degree d d d ; then a component of a field of maximal degree is not divisible by q q q . With P P P the 16 × r 16 \times r 16 × r matrix of fields and Λ \Lambda Λ the diagonal matrix of their eigenvalues,
P T P = q d I , P Λ P T = q d − 1 A , P Λ 2 P T = q d − 2 A 2 . P^{T}P=q^{d}I,\qquad P\Lambda P^{T}=q^{d-1}A,\qquad P\Lambda^{2}P^{T}=q^{d-2}A^{2}. P T P = q d I , P Λ P T = q d − 1 A , P Λ 2 P T = q d − 2 A 2 .
If d > 3 d>3 d > 3 , each diagonal entry of P Λ 2 P T P\Lambda^{2}P^{T} P Λ 2 P T is a sum of r ≤ 8 r \le8 r ≤ 8 squares divisible by q q q , and [24 Lemma 3.11] makes every field component divisible by q q q , a contradiction. So d = 1 d=1 d = 1 , the fields are J i x J_i x J i x with orthogonal skew, pairwise anticommuting J i J_i J i , and the Jacobi formula is Clifford. (This is the proof of [24 Lemma 3.3] with at most eight columns instead of six.) This proves Theorem F.
An extreme Jacobi eigenvalue of multiplicity seven or eight
Theorem G. Let R R R be an Osserman algebraic curvature tensor on R 16 \mathbb{R}^{16} R 16 with at least three distinct reduced Jacobi eigenvalues. If an eigenvalue of multiplicity seven or eight is the smallest or the largest reduced eigenvalue, then R R R is Clifford.
The hypothesis of three eigenvalues is needed: the Cayley tensor has two extreme eigenvalues. Throughout this section a reduced eigenvalue of multiplicity at least seven has been shifted to zero by subtracting a multiple of R 1 R_1 R 1 , so that the reduced Jacobi operator has nonzero rank ν ≤ 8 \nu\le8 ν ≤ 8 , and Λ \Lambda Λ is the diagonal matrix of the nonzero Jacobi eigenvalues for unit vectors, with multiplicities.
Proposition 9.1. Let R R R be an Osserman algebraic curvature tensor on R 16 \mathbb{R}^{16} R 16 as above.
If J x = M x Λ M x T J_x=M_x\Lambda M_x^{T} J x = M x Λ M x T has a global linear factor M x : R ν → R 16 M_x:\mathbb{R}^{\nu}\to\mathbb{R}^{16} M x : R ν → R 16 , injective for all x ≠ 0 x\ne0 x = 0 , then R R R is Clifford.
Suppose that J z = M z Λ M z T J_z=M_z\Lambda M_z^{T} J z = M z Λ M z T has a rational homogeneous factor of degree one with the following coherence property: there are a real unit vector X X X and an orthonormal eigenframe M X M_X M X such that, for all Y ⊥ X Y\perp X Y ⊥ X ,
M s X + Y = s M X + M Y (20) M_{sX+Y}=sM_X+M_Y \tag*{(20)} M s X + Y = s M X + M Y (20) as a rational identity. If the full Gram matrix Φ ( z ) = M z T M z \Phi(z)=M_z^{T}M_z Φ ( z ) = M z T M z on R 16 \mathbb{R}^{16} R 16 is polynomial, then R R R is Jacobi-proportional, hence, by Corollary D, Clifford or a linear combination of R 1 R_1 R 1 and an orthogonal conjugate of R C a y R_{\mathrm{Cay}} R Cay . If Λ \Lambda Λ is nonscalar, it is Clifford.
In part (2) the full Gram is Φ ( s X + Y ) = s 2 I + s ( M X T M Y + M Y T M X ) + M Y T M Y \Phi(sX+Y)=s^{2}I+s(M_X^{T}M_Y+M_Y^{T}M_X)+M_Y^{T}M_Y Φ ( s X + Y ) = s 2 I + s ( M X T M Y + M Y T M X ) + M Y T M Y , so its mixed coefficient must be polynomial as well.
Quadratic Gram rigidity
Lemma 9.2. Let Λ \Lambda Λ be invertible symmetric of size ν ≤ 8 \nu\le8 ν ≤ 8 . Suppose Φ ( x ) \Phi(x) Φ ( x ) is a symmetric ν × ν \nu\times\nu ν × ν matrix, homogeneous quadratic in x ∈ R 16 x \in\mathbb{R}^{16} x ∈ R 16 , positive definite for x ≠ 0 x \ne0 x = 0 , and Λ Φ ( x ) \Lambda\Phi(x) ΛΦ ( x ) has the spectrum of ∥ x ∥ 2 Λ \lVert x\rVert^2\Lambda ∥ x ∥ 2 Λ , with multiplicities. If Φ ( x 0 ) = I \Phi(x_0) = I Φ ( x 0 ) = I at a unit vector x 0 x_0 x 0 , then Φ ( x ) = ∥ x ∥ 2 I \Phi(x) = \lVert x\rVert^2I Φ ( x ) = ∥ x ∥ 2 I for every x x x .
Proof. This refines [23 Section 5, Lemmas 5–8, pp. 11–15 of arXiv:math/0204258v2] , whose hypothesis n > ( ν + 1 ) 2 / 4 n > (\nu+ 1)^2/4 n > ( ν + 1 ) 2 /4 fails for n = 16 n = 16 n = 16 , ν = 7 , 8 \nu= 7,8 ν = 7 , 8 ; the remaining pairs are settled by a topological count. Diagonalize Λ \Lambda Λ orthogonally, choose an eigenvalue λ \lambda λ of multiplicity m m m whose reciprocal is extremal, and set q = ∥ x ∥ 2 q = \lVert x\rVert^2 q = ∥ x ∥ 2 , Q λ ( x ) = λ q Λ − 1 − Φ ( x ) Q_\lambda(x) = \lambda q\Lambda^{-1} - \Phi(x) Q λ ( x ) = λ q Λ − 1 − Φ ( x ) . For unit x x x the matrix Φ ( x ) 1 / 2 Λ Φ ( x ) 1 / 2 \Phi(x)^{1/2}\Lambda\Phi(x)^{1/2} Φ ( x ) 1/2 ΛΦ ( x ) 1/2 is symmetric with the spectrum of Λ \Lambda Λ ; write it as O Λ O T O\Lambda O^T O Λ O T with O O O orthogonal. Then C = O T Φ ( x ) 1 / 2 C = O^T\Phi(x)^{1/2} C = O T Φ ( x ) 1/2 satisfies C Λ C T = Λ C\Lambda C^T = \Lambda C Λ C T = Λ , Φ = C T C \Phi= C^T C Φ = C T C and Q λ ( x ) = C T ( λ Λ − 1 − I ) C Q_\lambda(x) = C^T(\lambda\Lambda^{-1} - I)C Q λ ( x ) = C T ( λ Λ − 1 − I ) C , so Q λ ( x ) Q_\lambda(x) Q λ ( x ) has kernel dimension m m m for all x x x and is semidefinite of a fixed sign.
For u ∈ R ν u \in\mathbb{R}^{\nu} u ∈ R ν let H λ ( u ) H_\lambda(u) H λ ( u ) be the symmetric 16 × 16 16 \times16 16 × 16 matrix with x T H λ ( u ) x = u T Q λ ( x ) u x^T H_\lambda(u)x = u^TQ_\lambda(x)u x T H λ ( u ) x = u T Q λ ( x ) u ; it is quadratic in u u u and semidefinite, so L u : = { x : Q λ ( x ) u = 0 } = ker H λ ( u ) L_u := \{x : Q_\lambda(x)u = 0\} = \ker H_\lambda(u) L u := { x : Q λ ( x ) u = 0 } = ker H λ ( u ) is a linear subspace, of dimension d ( u ) d(u) d ( u ) . Choose independent u 1 , … , u m u_1,\ldots,u_m u 1 , … , u m maximizing ∑ i d ( u i ) \sum_i d(u_i) ∑ i d ( u i ) and put S = span ( u 1 , … , u m ) S = \operatorname{span}(u_1,\ldots,u_m) S = span ( u 1 , … , u m ) . For u ∉ S u \notin S u ∈ / S and each i i i , the vectors u u u and u j u_j u j (j ≠ i j \ne i j = i ) are independent, so d ( u ) ≤ d ( u i ) d(u) \le d(u_i) d ( u ) ≤ d ( u i ) . For any m + 1 m+1 m + 1 independent coefficient vectors v 0 , … , v m v_0,\ldots,v_m v 0 , … , v m one has L v 0 ∩ ⋯ ∩ L v m = 0 L_{v_0} \cap\cdots\cap L_{v_m} = 0 L v 0 ∩ ⋯ ∩ L v m = 0 (otherwise ker Q λ ( x ) \ker Q_\lambda(x) ker Q λ ( x ) would have dimension at least m + 1 m+1 m + 1 at some x ≠ 0 x \ne0 x = 0 ), and the subspace intersection inequality 0 = dim ( L u ∩ L u 1 ∩ ⋯ ∩ L u m ) ≥ d ( u ) + ∑ i d ( u i ) − 16 m 0 = \dim(L_u \cap L_{u_1} \cap\cdots\cap L_{u_m}) \ge d(u) + \sum_i d(u_i) - 16m 0 = dim ( L u ∩ L u 1 ∩ ⋯ ∩ L u m ) ≥ d ( u ) + ∑ i d ( u i ) − 16 m , with d ( u ) ≤ d ( u i ) d(u) \le d(u_i) d ( u ) ≤ d ( u i ) for each i i i , gives ( m + 1 ) d ( u ) ≤ d ( u ) + ∑ i d ( u i ) ≤ 16 m (m+1)d(u) \le d(u) + \sum_i d(u_i) \le16m ( m + 1 ) d ( u ) ≤ d ( u ) + ∑ i d ( u i ) ≤ 16 m , that is,
d ( u ) ≤ h : = ⌊ 16 m m + 1 ⌋ ( u ∉ S ) . (21) d(u) \le h := \left\lfloor\frac{16m}{m+1} \right\rfloor\qquad(u \notin S). \tag*{(21)} d ( u ) ≤ h := ⌊ m + 1 16 m ⌋ ( u ∈ / S ) . (21)
The incidence space I = { ( x , u ) : x ≠ 0 , Q λ ( x ) u = 0 } \mathcal{I} = \{(x,u) : x \ne0, Q_\lambda(x)u = 0\} I = {( x , u ) : x = 0 , Q λ ( x ) u = 0 } is a connected real-analytic rank-m m m vector bundle over R 16 ∖ { 0 } \mathbb{R}^{16} \setminus\{0\} R 16 ∖ { 0 } , of dimension 16 + m 16+m 16 + m , and its open subset I o u t \mathcal{I}_{\mathrm{out}} I out where u ∉ S u \notin S u ∈ / S has semialgebraic dimension at most ν + h \nu+h ν + h . If ν + h < 16 + m \nu+h < 16+m ν + h < 16 + m , then I o u t \mathcal{I}_{\mathrm{out}} I out is empty and ker Q λ ( x ) = S \ker Q_\lambda(x) = S ker Q λ ( x ) = S for every x x x . The scalar case m = ν m=\nu m = ν is immediate, and for ν ≤ 8 \nu\le8 ν ≤ 8 the only remaining nonscalar pairs are
Table 1 (PDF p. 16).
Suppose I o u t ≠ ∅ \mathcal{I}_{\mathrm{out}} \ne\varnothing I out = ∅ in one of these cases and put ℓ = 16 + m − ν \ell= 16+m-\nu ℓ = 16 + m − ν . The coefficient strata with d ( u ) ≤ ℓ − 1 d(u) \le\ell-1 d ( u ) ≤ ℓ − 1 have incidence dimension at most ν + ℓ − 1 = 15 + m \nu+\ell-1 = 15+m ν + ℓ − 1 = 15 + m , so the locus with d ( u ) ≥ ℓ d(u) \ge\ell d ( u ) ≥ ℓ contains a relatively open subset of I \mathcal{I} I . All ( 17 − ℓ ) (17-\ell) ( 17 − ℓ ) -minors of H λ ( u ) H_\lambda(u) H λ ( u ) vanish there; their pullbacks to I \mathcal{I} I are real analytic, so by connectedness they vanish on I \mathcal{I} I , and every nonempty coefficient fibre has d ( u ) ≥ ℓ d(u) \ge\ell d ( u ) ≥ ℓ . Normalize both variables:
E = { ( x , u ) ∈ S 15 × S ν − 1 : Q λ ( x ) u = 0 } , Z = { u ∈ S ν − 1 : det H λ ( u ) = 0 } . \mathcal{E} = \{(x,u) \in S^{15} \times S^{\nu-1} : Q_\lambda(x)u = 0\}, \qquad Z = \{u \in S^{\nu-1} : \det H_\lambda(u) = 0\}. E = {( x , u ) ∈ S 15 × S ν − 1 : Q λ ( x ) u = 0 } , Z = { u ∈ S ν − 1 : det H λ ( u ) = 0 } .
The first projection makes E \mathcal{E} E an S m − 1 S^{m-1} S m − 1 -bundle over S 15 S^{15} S 15 , so dim E = 14 + m \dim\mathcal{E} = 14+m dim E = 14 + m . The second projection is a proper surjection onto the compact real algebraic set Z Z Z , with fibres S ( L u ) S(L_u) S ( L u ) of dimension at least ℓ − 1 \ell-1 ℓ − 1 , and at most h − 1 h-1 h − 1 outside S S S by (9.2). As E o u t \mathcal{E}_{\mathrm{out}} E out is nonempty and open,
15 + m − h ≤ z : = dim Z ≤ 15 + m − ℓ = ν − 1 , 15+m-h \le z := \dim Z \le15+m-\ell= \nu-1, 15 + m − h ≤ z := dim Z ≤ 15 + m − ℓ = ν − 1 ,
which gives
Table 4.
In every row z < ℓ − 1 z < \ell- 1 z < ℓ − 1 , so the finite-degree Vietoris–Begle theorem [19 Theorem 1.1 (Begle), p. 383] , with mod-two coefficients, gives H z ( Z ; F 2 ) ≅ H z ( E ; F 2 ) H^z(Z;\mathbb{F}_2) \cong H^z(E;\mathbb{F}_2) H z ( Z ; F 2 ) ≅ H z ( E ; F 2 ) (all spaces are compact semialgebraic and triangulable, so Čech and singular cohomology agree). The latter group vanishes by the Gysin sequence of the S m − 1 S^{m-1} S m − 1 -bundle over S 15 S^{15} S 15 : the base groups in degrees z z z and z − m + 1 z-m+1 z − m + 1 vanish in every row. But a nonempty compact real algebraic set of dimension z z z has a nonzero top mod-two homology class [5 Theorem 1.1, p. 22] , hence H z ( Z ; F 2 ) ≠ 0 H_z(Z;\mathbb{F}_2) \ne0 H z ( Z ; F 2 ) = 0 , a contradiction.
Hence ker Q λ ( x ) = S \ker Q_\lambda(x)=S ker Q λ ( x ) = S for all x ≠ 0 x\ne0 x = 0 ; at x 0 x_0 x 0 this is the λ \lambda λ -eigenspace of Λ \Lambda Λ , so Φ ( x ) u = λ q Λ − 1 u = q u \Phi(x)u=\lambda q\Lambda^{-1}u=qu Φ ( x ) u = λ q Λ − 1 u = q u for u ∈ S u\in S u ∈ S . Symmetry removes the cross blocks, and restricting Φ , Λ \Phi,\Lambda Φ , Λ to the orthogonal complement preserves all hypotheses. Induction on ν \nu ν proves Φ = q I \Phi=qI Φ = q I .
Proof of Proposition 9.1
Proof of part (1). A constant coefficient change M x ↦ M x G 0 M_x\mapsto M_xG_0 M x ↦ M x G 0 normalizes M x 0 T M x 0 = I M_{x_0}^{T}M_{x_0}=I M x 0 T M x 0 = I at one unit vector and replaces Λ \Lambda Λ by G 0 − 1 Λ G 0 − T G_0^{-1}\Lambda G_0^{-T} G 0 − 1 Λ G 0 − T ; this replaces Λ Φ ( x ) \Lambda\Phi(x) ΛΦ ( x ) by the similar matrix G 0 − 1 Λ Φ ( x ) G 0 G_0^{-1}\Lambda\Phi(x)G_0 G 0 − 1 ΛΦ ( x ) G 0 , and at x 0 x_0 x 0 the new Λ \Lambda Λ has the spectrum of the old one, so the hypotheses of Lemma 9.2 are kept. For a global linear factor, Φ ( x ) = M x T M x \Phi(x)=M_x^{T}M_x Φ ( x ) = M x T M x is homogeneous quadratic and positive definite, and the nonzero spectrum of J x J_x J x is that of Λ Φ ( x ) \Lambda\Phi(x) ΛΦ ( x ) , so Lemma 9.2 gives M x T M x = q I M_x^{T}M_x=qI M x T M x = q I . For a coefficient basis vector e i e_i e i write J i x = M x e i J_i x=M_xe_i J i x = M x e i . Since J x x = 0 J_xx=0 J x x = 0 , M x M_x M x is injective and Λ \Lambda Λ is invertible, M x T x = 0 M_x^{T}x=0 M x T x = 0 , so each J i J_i J i is skew; the Gram identity makes the J i J_i J i orthogonal with J i J j + J j J i = − 2 δ i j I J_iJ_j+J_jJ_i=-2\delta_{ij}I J i J j + J j J i = − 2 δ ij I by polarization. Diagonalizing Λ \Lambda Λ gives the Clifford formula (with no columns, R R R has constant curvature).
Proof of part (2). On the real regular locus Φ = M T M \Phi=M^{T}M Φ = M T M is positive definite, and the nonzero spectrum gives det Φ ( z ) = q ( z ) ν \det\Phi(z)=q(z)^\nu det Φ ( z ) = q ( z ) ν . If Φ \Phi Φ is polynomial, it is positive semidefinite everywhere by continuity and positive definite for z ≠ 0 z\ne0 z = 0 by the determinant. By (9.1) and homogeneity, M X + t Y = M X + t M Y M_{X+tY}=M_X+tM_Y M X + t Y = M X + t M Y for generic Y ⊥ X Y\perp X Y ⊥ X and t ≠ 0 t\ne0 t = 0 , and continuity at t = 0 t=0 t = 0 gives Φ ( X ) = I \Phi(X)=I Φ ( X ) = I . Lemma 9.2 yields Φ ( z ) = q ( z ) I \Phi(z)=q(z)I Φ ( z ) = q ( z ) I .
For a nonzero eigenvalue λ j \lambda_j λ j let I j I_j I j be its coefficient projector and π j ( z ) \pi_j(z) π j ( z ) the corresponding reduced Jacobi projector. Then C j ( z ) : = M z I j M z T = q ( z ) π j ( z ) C_j(z):=M_zI_jM_z^{T}=q(z)\pi_j(z) C j ( z ) := M z I j M z T = q ( z ) π j ( z ) on the regular locus, hence as a rational identity. In coordinates z = s X + Y z=sX+Y z = s X + Y the left side has a denominator dividing a power of a polynomial D ( Y ) D(Y) D ( Y ) independent of s s s , by (9.1), while the right side has only powers of q ( z ) = s 2 + ∥ Y ∥ 2 q(z)=s^2+\lVert Y\rVert^2 q ( z ) = s 2 + ∥ Y ∥ 2 as possible denominators: by Lagrange interpolation over the distinct nonzero eigenvalues λ 1 , … , λ ρ \lambda_1,\ldots,\lambda_\rho λ 1 , … , λ ρ , π j ( z ) = J z ∏ k ≠ j ( J z − λ k q ( z ) I ) / ( λ j ∏ k ≠ j ( λ j − λ k ) q ( z ) ρ ) \pi_j(z)=J_z\prod_{k\ne j}(J_z-\lambda_kq(z)I)/(\lambda_j\prod_{k\ne j}(\lambda_j-\lambda_k)q(z)^\rho) π j ( z ) = J z ∏ k = j ( J z − λ k q ( z ) I ) / ( λ j ∏ k = j ( λ j − λ k ) q ( z ) ρ ) . The polynomial q ( z ) q(z) q ( z ) is irreducible over C \mathbb{C} C in sixteen variables and coprime to every nonzero D ( Y ) D(Y) D ( Y ) , so unique factorization shows that C j C_j C j is polynomial, and homogeneity makes it quadratic; so is the reduced zero-eigenvalue family C 0 ( z ) = q ( z ) I − z z T − ∑ j C j ( z ) C_0(z)=q(z)I-zz^T-\sum_jC_j(z) C 0 ( z ) = q ( z ) I − z z T − ∑ j C j ( z ) . Each C j C_j C j is symmetric with C j ( z ) z = 0 C_j(z)z=0 C j ( z ) z = 0 , and the polarization argument of Section 2 gives ⟨ C j ( x ) y , y ⟩ = ⟨ C j ( y ) x , x ⟩ \langle C_j(x)y,y\rangle=\langle C_j(y)x,x\rangle ⟨ C j ( x ) y , y ⟩ = ⟨ C j ( y ) x , x ⟩ . So R R R is Jacobi-proportional, and Corollary D applies. If Λ \Lambda Λ is nonscalar, the zero eigenvalue and at least two distinct nonzero eigenvalues give three reduced eigenvalues, which excludes the Cayley alternative.
The coherent partner
Lemma 9.3. For the shifted tensor above, with ν ≤ 8 \nu\le8 ν ≤ 8 , there are a real unit anchor X X X , an orthonormal eigenframe M X M_X M X , and a rational homogeneous partner M Y M_Y M Y on X ⊥ X^\perp X ⊥ with
J s X + Y = ( s M X + M Y ) Λ ( s M X + M Y ) T . \mathcal{J}_{sX+Y}=(sM_X+M_Y)\Lambda(sM_X+M_Y)^T. J s X + Y = ( s M X + M Y ) Λ ( s M X + M Y ) T .
Proof. We use two results of [23 Section 4] , both stated for n ≥ 2 ν n \ge2\nu n ≥ 2 ν , hence valid for n = 16 n=16 n = 16 , ν ≤ 8 \nu\le8 ν ≤ 8 . By [23 Lemma 2(1) and its proof, p. 7 of arXiv:math/0204258v2] , for every unit X X X the unit vectors Y ⊥ X Y \perp X Y ⊥ X with Im J X ∩ Im J Y = 0 \operatorname{Im}\mathcal{J}_X \cap\operatorname{Im}\mathcal{J}_Y=0 Im J X ∩ Im J Y = 0 form an open dense subset of the unit sphere of X ⊥ X^\perp X ⊥ . By [23 Lemma 3, p. 8 of arXiv:math/0204258v2] , for every such orthonormal pair ( X , Y ) (X,Y) ( X , Y ) and every M X M_X M X with J X = M X Λ M X T \mathcal{J}_X=M_X\Lambda M_X^T J X = M X Λ M X T there is a unique M Y M_Y M Y with J s X + t Y = ( s M X + t M Y ) Λ ( s M X + t M Y ) T \mathcal{J}_{sX+tY}=(sM_X+tM_Y)\Lambda(sM_X+tM_Y)^T J s X + t Y = ( s M X + t M Y ) Λ ( s M X + t M Y ) T for all real s , t s,t s , t . Fix a unit X X X and an orthonormal eigenframe M X M_X M X , and put M c Y = c M Y M_{cY}=cM_Y M c Y = c M Y for c ≠ 0 c\ne0 c = 0 ; uniqueness makes this consistent. Write M Y = M X F ( Y ) + K X P ( Y ) M_Y=M_XF(Y)+K_XP(Y) M Y = M X F ( Y ) + K X P ( Y ) , where K X K_X K X is an orthonormal frame of ker J X \ker\mathcal{J}_X ker J X , and 2 B ( X , Y ) = J X + Y − J X − J Y 2B(X,Y)=\mathcal{J}_{X+Y}-\mathcal{J}_X-\mathcal{J}_Y 2 B ( X , Y ) = J X + Y − J X − J Y . The mixed coefficient gives P ( Y ) = K X T 2 B ( X , Y ) M X Λ − 1 P(Y)=K_X^T2B(X,Y)M_X\Lambda^{-1} P ( Y ) = K X T 2 B ( X , Y ) M X Λ − 1 , which is linear, of full column rank on the generic-pair locus since the images of J X \mathcal{J}_X J X and J Y \mathcal{J}_Y J Y meet only in 0 0 0 . The lower-upper block of J Y \mathcal{J}_Y J Y gives P Λ F T = K X T J Y M X P\Lambda F^T=K_X^T\mathcal{J}_YM_X P Λ F T = K X T J Y M X , and the rational left inverse ( P T P ) − 1 P T (P^TP)^{-1}P^T ( P T P ) − 1 P T shows that F F F , hence M Y M_Y M Y , is rational homogeneous of degree one. The two-plane identities hold rationally by continuation, and defining M s X + Y = s M X + M Y M_{sX+Y}=sM_X+M_Y M s X + Y = s M X + M Y gives the required coherent factor.
Proof of Theorem G
Proof. Shift the extreme eigenvalue of multiplicity seven or eight to zero; then ν = 7 \nu=7 ν = 7 or 8 8 8 and Λ \Lambda Λ is definite. Use Lemma 9.3 with coordinates Y = ( y 1 , … , y 15 ) Y=(y_1,\ldots,y_{15}) Y = ( y 1 , … , y 15 ) on X ⊥ X^\perp X ⊥ , and put F = R ( y 1 , … , y 15 ) \mathbb{F}=\mathbb{R}(y_1,\ldots,y_{15}) F = R ( y 1 , … , y 15 ) , d = ∑ i y i 2 d=\sum_i y_i^2 d = ∑ i y i 2 , G ( s ) = s 2 I + s Φ 1 + Φ 0 G(s)=s^2I+s\Phi_1+\Phi_0 G ( s ) = s 2 I + s Φ 1 + Φ 0 , where Φ 1 = M X T M Y + M Y T M X \Phi_1=M_X^TM_Y+M_Y^TM_X Φ 1 = M X T M Y + M Y T M X and Φ 0 = M Y T M Y \Phi_0=M_Y^TM_Y Φ 0 = M Y T M Y have entries in F \mathbb{F} F ; thus G ( s ) = Φ ( s X + Y ) G(s)=\Phi(sX+Y) G ( s ) = Φ ( s X + Y ) is the full Gram of Proposition 9.1 . Equality of the nonzero Jacobi spectra gives
det ( t I − Λ G ( s ) ) = ∏ i = 1 ν ( t − ( s 2 + d ) μ i ) in F [ s , t ] , (22) \det(tI-\Lambda G(s))=\prod_{i=1}^{\nu}\bigl(t-(s^2+d)\mu_i\bigr)\quad\text{in }\mathbb{F}[s,t], \tag*{(22)} det ( t I − Λ G ( s )) = i = 1 ∏ ν ( t − ( s 2 + d ) μ i ) in F [ s , t ] , (22)
where the μ i \mu_i μ i are the nonzero eigenvalues with multiplicities.
Let E = F ( α ) \mathbb{E}=\mathbb{F}(\alpha) E = F ( α ) with α 2 = − d \alpha^2=-d α 2 = − d ; as F \mathbb{F} F is formally real, [ E : F ] = 2 [\mathbb{E}:\mathbb{F}]=2 [ E : F ] = 2 , and E \mathbb{E} E is the fraction field of R [ s , y 1 , … , y 15 ] / ( s 2 + y 1 2 + ⋯ + y 15 2 ) \mathbb{R}[s,y_1,\ldots,y_{15}]/(s^2+y_1^2+\cdots+y_{15}^2) R [ s , y 1 , … , y 15 ] / ( s 2 + y 1 2 + ⋯ + y 15 2 ) . By [24 Lemma 3.11 and its proof, p. 398] , a sum of at most eight squares in E \mathbb{E} E vanishes only when every term vanishes, so every constant definite real quadratic form of dimension at most eight is anisotropic over E \mathbb{E} E . At s = α s=\alpha s = α , (22) shows that N = Λ G ( α ) N=\Lambda G(\alpha) N = Λ G ( α ) is nilpotent, and it is self-adjoint for the anisotropic form ⟨ v , w ⟩ Λ − 1 = v T Λ − 1 w \langle v,w\rangle_{\Lambda^{-1}}=v^T\Lambda^{-1}w ⟨ v , w ⟩ Λ − 1 = v T Λ − 1 w , since Λ − 1 N = G ( α ) \Lambda^{-1}N=G(\alpha) Λ − 1 N = G ( α ) is symmetric. Such an N N N is zero: if its nilpotency index were e ≥ 2 e\ge2 e ≥ 2 , then w = N e − 1 v ≠ 0 w=N^{e-1}v\ne0 w = N e − 1 v = 0 for some v v v and ⟨ w , w ⟩ Λ − 1 = ⟨ v , N 2 e − 2 v ⟩ Λ − 1 = 0 \langle w,w\rangle_{\Lambda^{-1}}=\langle v,N^{2e-2}v\rangle_{\Lambda^{-1}}=0 ⟨ w , w ⟩ Λ − 1 = ⟨ v , N 2 e − 2 v ⟩ Λ − 1 = 0 , contradicting anisotropy. As Λ \Lambda Λ is invertible, Φ 0 − d I + α Φ 1 = 0 \Phi_0-dI+\alpha\Phi_1=0 Φ 0 − d I + α Φ 1 = 0 , and the independence of 1 , α 1,\alpha 1 , α over F \mathbb{F} F gives Φ 1 = 0 \Phi_1=0 Φ 1 = 0 and Φ 0 = d I \Phi_0=dI Φ 0 = d I . So the full Gram is G ( s ) = ( s 2 + d ) I G(s)=(s^2+d)I G ( s ) = ( s 2 + d ) I , polynomial on R 16 \mathbb{R}^{16} R 16 , and Proposition 9.1 (2) makes R R R Clifford (there are at least three reduced eigenvalues).
The coherent factor and its transpose pencil
From here to the end of Section 26, R R R is an Osserman algebraic curvature tensor on R 16 \mathbb{R}^{16} R 16 with a reduced eigenvalue μ \mu μ of multiplicity seven or eight that is neither the smallest nor the largest reduced eigenvalue (the extreme case is Theorem G). Shift μ \mu μ to zero; the shifted Jacobi operators then have rank ν ∈ { 7 , 8 } \nu\in\{7,8\} ν ∈ { 7 , 8 } ; put k = 16 − ν k=16-\nu k = 16 − ν . Let Λ \Lambda Λ be the diagonal matrix of the nonzero eigenvalues, with multiplicities; it is indefinite because μ \mu μ is neither the smallest nor the largest reduced eigenvalue. Put H = Λ − 1 H=\Lambda^{-1} H = Λ − 1 and B ( X , Y ) = 1 2 ( J X + Y − J X − J Y ) B(X,Y)=\frac{1}{2}(J_{X+Y}-J_X-J_Y) B ( X , Y ) = 2 1 ( J X + Y − J X − J Y ) . For a real unit anchor X X X choose orthonormal frames M X : R ν → E X = Im J X M_X:\mathbb{R}^{\nu}\to E_X=\operatorname{Im}J_X M X : R ν → E X = Im J X and K X : R k → ker J X K_X:\mathbb{R}^{k}\to\ker J_X K X : R k → ker J X , with J X = M X Λ M X T J_X=M_X\Lambda M_X^{T} J X = M X Λ M X T . A regular anchor is a point of the nonempty Euclidean-open set U 0 U_0 U 0 constructed in § 27 : the generic-pair construction below applies there, and U 0 U_0 U 0 avoids the proper algebraic exceptional loci named in the following sections. § 27 partitions U 0 U_0 U 0 into finitely many semialgebraic sets U j U_j U j on which all pole invariants are constant (Lemma 27.1 ); we call them branches . It also lists the cases, (L1)–(L8), and names the section that settles each; a first reading may start there.
Construction and normalization
Take X X X in the nonempty open set of anchors where the generic-pair construction of Lemma 9.3 applies (only the two-plane result of [23 Section 4] is used, not its three-plane construction, whose dimension hypothesis fails here). As in that proof, M Y = M X F ( Y ) + K X P ( Y ) M_Y=M_XF(Y)+K_XP(Y) M Y = M X F ( Y ) + K X P ( Y ) with
P ( Y ) = K X T 2 B ( X , Y ) M X Λ − 1 (23) P(Y)=K_X^{T}2B(X,Y)M_X\Lambda^{-1} \tag*{(23)} P ( Y ) = K X T 2 B ( X , Y ) M X Λ − 1 (23)
linear, F F F rational and homogeneous of degree one, and the block identities
F Λ + Λ F T = M X T 2 B ( X , Y ) M X , (24) F\Lambda+\Lambda F^{T}=M_X^{T}2B(X,Y)M_X, \tag*{(24)} F Λ + Λ F T = M X T 2 B ( X , Y ) M X , (24)
P Λ F T = K X T J Y M X , (25) P\Lambda F^{T}=K_X^{T}J_YM_X, \tag*{(25)} P Λ F T = K X T J Y M X , (25)
F Λ F T = M X T J Y M X , P Λ P T = K X T J Y K X (26) F\Lambda F^{T}=M_X^{T}J_YM_X,\qquad P\Lambda P^{T}=K_X^{T}J_YK_X \tag*{(26)} F Λ F T = M X T J Y M X , P Λ P T = K X T J Y K X (26)
hold as rational identities; P P P , which we call the lower pencil , is injective for generic Y Y Y . Thus
M s X + Y = s M X + M X F ( Y ) + K X P ( Y ) , J s X + Y = M s X + Y Λ M s X + Y T . (27) M_{sX+Y}=sM_X+M_XF(Y)+K_XP(Y),\qquad J_{sX+Y}=M_{sX+Y}\Lambda M_{sX+Y}^{T}. \tag*{(27)} M s X + Y = s M X + M X F ( Y ) + K X P ( Y ) , J s X + Y = M s X + Y Λ M s X + Y T . (27)
Lemma 10.1. If F F F has no pole along any complex hypersurface in X ⊥ X^\perp X ⊥ , then (10.5) is a global linear factor, and the tensor is Clifford.
Proof. In the polynomial ring on X ⊥ X^\perp X ⊥ , a rational function with no divisorial pole is polynomial; by homogeneity F F F is linear, and (10.5) is a polynomial identity in all sixteen variables. Its factor is injective at every nonzero real vector, as the Jacobi operator has rank ν \nu ν , so Proposition 9.1 (1) applies.
Kernel directions give Jacobi factors
Extend P P P linearly to V = R 16 V=\mathbb{R}^{16} V = R 16 by P ( X ) = 0 P(X)=0 P ( X ) = 0 . For g ∈ R k g\in\mathbb{R}^{k} g ∈ R k define N g : R ν → V N_g:\mathbb{R}^{\nu}\to V N g : R ν → V by
N g T Y = P ( Y ) T g ( Y ∈ V ) , (28) N_g^{T}Y=P(Y)^{T}g\qquad(Y\in V), \tag*{(28)} N g T Y = P ( Y ) T g ( Y ∈ V ) , (28)
linear in g g g , and for a ∈ R ν a\in\mathbb{R}^{\nu} a ∈ R ν define the transpose pencil B ( a ) : R k → V B(a):\mathbb{R}^{k}\to V B ( a ) : R k → V by B ( a ) g = N g a B(a)g=N_ga B ( a ) g = N g a .
Lemma 10.2. For every real g g g ,
J K X g = N g Λ N g T . (29) J_{K_Xg}=N_g\Lambda N_g^{T}. \tag*{(29)} J K X g = N g Λ N g T . (29) If g ≠ 0 g\ne0 g = 0 , then N g N_g N g is injective. If a ≠ 0 a\ne0 a = 0 is real, then B ( a ) B(a) B ( a ) is injective; equivalently Y ↦ P ( Y ) a Y\mapsto P(Y)a Y ↦ P ( Y ) a is onto R k \mathbb{R}^{k} R k . There are no nonzero fixed real coefficient space U U U and proper fixed real subspace T ⊂ R k T\subset\mathbb{R}^{k} T ⊂ R k with P ( Y ) U ⊂ T P(Y)U\subset T P ( Y ) U ⊂ T for all Y Y Y .
Proof. For z = K X g z = K_Xg z = K X g and v = s X + Y v = sX + Y v = s X + Y with Y ⊥ X Y \perp X Y ⊥ X , the relations M X T z = 0 M_X^Tz = 0 M X T z = 0 and P ( X ) = 0 P(X) = 0 P ( X ) = 0 give M X T z = P ( Y ) T g = N g T v M_X^Tz = P(Y)^Tg = N_g^Tv M X T z = P ( Y ) T g = N g T v for every real s s s . So the coherent factor and the symmetry ⟨ J v z , z ⟩ = ⟨ J z v , v ⟩ \langle J_{v z},z\rangle= \langle J_{z v},v\rangle ⟨ J v z , z ⟩ = ⟨ J z v , v ⟩ give
⟨ J z v , v ⟩ = ( N g T v ) T Λ ( N g T v ) ( v ∈ R 16 ) . \langle J_{z v},v\rangle= (N_g^Tv)^T\Lambda(N_g^Tv) \qquad(v \in\mathbb{R}^{16}). ⟨ J z v , v ⟩ = ( N g T v ) T Λ ( N g T v ) ( v ∈ R 16 ) .
polarizing in v v v proves (10.7). Its left side has rank ν \nu ν for real g ≠ 0 g \ne0 g = 0 , so N g N_g N g is injective, and N g a = 0 N_ga = 0 N g a = 0 with a ≠ 0 a \ne0 a = 0 forces g = 0 g = 0 g = 0 ; transposition gives surjectivity. Finally, for nonzero g ⊥ T g \perp T g ⊥ T the containment would make N g N_g N g vanish on U U U .
Let g X g_X g X be the unit vector with K X g X = X K_Xg_X = X K X g X = X . Then
B ( a ) g X = − M X a , K X T B ( a ) is skew-symmetric , Im B ( a ) ⊂ X ⊥ . (30) B(a)g_X = -M_Xa,\qquad K_X^TB(a)\text{ is skew-symmetric},\qquad\operatorname{Im} B(a) \subset X^\perp. \tag*{(30)} B ( a ) g X = − M X a , K X T B ( a ) is skew-symmetric , Im B ( a ) ⊂ X ⊥ . (30)
Indeed, injectivity of the coherent factor and J z z = 0 J_{z z} = 0 J z z = 0 give M z T z = 0 M_z^Tz = 0 M z T z = 0 ; the coefficient of s s s for z = s X + Y z = sX + Y z = s X + Y gives P ( Y ) T g X = − M X T Y P(Y)^Tg_X = -M_X^TY P ( Y ) T g X = − M X T Y , the first identity. The last follows from P ( X ) = 0 P(X) = 0 P ( X ) = 0 . For the middle one, (10.7) and J K X g K X g = 0 J_{K_Xg\,K_Xg} = 0 J K X g K X g = 0 give N g T K X g = 0 N_g^TK_Xg = 0 N g T K X g = 0 , so ⟨ g , K X T B ( a ) g ⟩ = 0 \langle g,K_X^TB(a)g\rangle= 0 ⟨ g , K X T B ( a ) g ⟩ = 0 for all real g , a g,a g , a ; polarize in g g g . We call the first identity in (10.8), in either form, the radial identity ; B ( a ) g X = − M X a B(a)g_X = -M_Xa B ( a ) g X = − M X a is the radial column and g X T P ( Y ) = − Y T M X g_X^TP(Y) = -Y^TM_X g X T P ( Y ) = − Y T M X the radial row . For ν = 8 \nu= 8 ν = 8 (so k = ν k = \nu k = ν ) we say that the transpose pencil is aligned at X X X if Im B ( a ) = E X \operatorname{Im} B(a) = E_X Im B ( a ) = E X for every real a ≠ 0 a \ne0 a = 0 .
Lemma 10.3. For every nonzero complex coefficient a a a , the kernel of B ( a ) B(a) B ( a ) is totally isotropic in C k \mathbb{C}^k C k for the standard bilinear form, and s ( a ) = rank B ( a ) s(a) = \operatorname{rank} B(a) s ( a ) = rank B ( a ) satisfies s ( a ) ≥ ⌈ k / 2 ⌉ s(a) \ge\lceil k/2\rceil s ( a ) ≥ ⌈ k /2 ⌉ and $\operatorname{rank}(B(a)B(a)^T) = 2s(a) - k.
Proof. If 0 ≠ g ∈ ker B ( a ) 0 \ne g \in\ker B(a) 0 = g ∈ ker B ( a ) with g T g ≠ 0 g^Tg \ne0 g T g = 0 , the Osserman characteristic and minimal-polynomial identities, continued to complex g g g , make J K X g J_{K_Xg} J K X g diagonalizable of rank ν \nu ν , so by (10.7), continued polynomially in g g g , N g N_g N g has rank ν \nu ν , contrary to N g a = 0 N_ga = 0 N g a = 0 . So L = ker B ( a ) L = \ker B(a) L = ker B ( a ) consists of null vectors, is totally isotropic by polarization, and has dimension at most ⌊ k / 2 ⌋ \lfloor k/2\rfloor ⌊ k /2 ⌋ . As Im B ( a ) T = L ⊥ ⊃ L \operatorname{Im} B(a)^T = L^\perp\supset L Im B ( a ) T = L ⊥ ⊃ L , the image of B ( a ) B ( a ) T B(a)B(a)^T B ( a ) B ( a ) T has dimension dim L ⊥ − dim L = 2 s ( a ) − k \dim L^\perp- \dim L = 2s(a) - k dim L ⊥ − dim L = 2 s ( a ) − k .
The local pole budget
Put F 0 = M X T B ( X , Y ) M X Λ − 1 F_0 = M_X^TB(X,Y)M_X\Lambda^{-1} F 0 = M X T B ( X , Y ) M X Λ − 1 , linear in Y Y Y , and Ω = ( F − F 0 ) Λ \Omega= (F - F_0)\Lambda Ω = ( F − F 0 ) Λ , a skew-symmetric matrix, homogeneous of degree one. Since Ω \Omega Ω is skew and Λ \Lambda Λ symmetric, P Ω = − P Ω T = − P Λ F T + P Λ F 0 T P\Omega= -P\Omega^T = -P\Lambda F^T + P\Lambda F_0^T P Ω = − P Ω T = − P Λ F T + P Λ F 0 T , which is polynomial by (10.3), as P P P and F 0 F_0 F 0 are linear; and, as F = F 0 + Ω H F = F_0 + \Omega H F = F 0 + Ω H and H Λ = I H\Lambda= I H Λ = I , the first identity in (10.4) makes Ω H Ω T + Ω F 0 T + F 0 Ω T = F Λ F T − F 0 Λ F 0 T \Omega H\Omega^T + \Omega F_0^T + F_0\Omega^T = F\Lambda F^T - F_0\Lambda F_0^T Ω H Ω T + Ω F 0 T + F 0 Ω T = F Λ F T − F 0 Λ F 0 T polynomial; so the negative-order terms of Ω H Ω T \Omega H\Omega^T Ω H Ω T are cancelled only by terms whose pole order is at most that of Ω \Omega Ω .
Lemma 10.4. Let f f f be an irreducible polynomial, real or complex, defining a pole of Ω \Omega Ω of order d > 0 d > 0 d > 0 . Every maximal ν \nu ν -minor of P P P is divisible by f 2 d f^{2d} f 2 d ; if d = 1 d = 1 d = 1 and the residue has rank r r r , every such minor is divisible by f r f^r f r . Let the reduced common denominator D D D of Ω \Omega Ω be the least common multiple of the denominators of its entries, each entry written in lowest terms; a pole of order d d d along f f f contributes f d f^d f d , so D D D need not be squarefree. Then D 2 D^2 D 2 divides every maximal minor of P P P , and 2 deg D ≤ ν 2\deg D \le\nu 2 deg D ≤ ν .
Proof. Localize the polynomial ring at ( f ) (f) ( f ) , a discrete valuation ring with uniformizer π = f \pi= f π = f . An alternating matrix over its fraction field is unimodularly congruent to two-by-two blocks a j ( 0 1 − 1 0 ) a_j\begin{pmatrix}0&1\\-1&0\end{pmatrix} a j ( 0 − 1 1 0 ) and a zero block (pivot on an entry of least valuation and clear its two rows and columns; all coefficients are integral). Let the negative block valuations be − d j -d_j − d j . As P Ω P\Omega P Ω is integral, the corresponding two columns of the transformed P P P are divisible by π d j \pi^{d_j} π d j , so every maximal minor has valuation at least 2 ∑ j d j ≥ 2 d 2\sum_j d_j \ge2d 2 ∑ j d j ≥ 2 d , since the largest d j d_j d j is the pole order d d d . At a simple pole the residue rank is twice the number of negative blocks. Unimodular column changes preserve the maximal-minor ideal, and distinct irreducible factors multiply, which gives the claim for D D D . Some maximal minor is nonzero, of degree ν \nu ν , since P P P is generically injective.
If Ω \Omega Ω has a pole of order d > 0 d>0 d > 0 along an irreducible f f f , call W = ( f d Ω ) mod f W=(f^d\Omega)\mathbin{\operatorname{mod}} f W = ( f d Ω ) mod f , a matrix over the residue field of ( f ) (f) ( f ) , the leading Laurent coefficient of Ω \Omega Ω at f f f . The leading terms of the same identities give, over the residue field,
P W = 0 , W T = − W , W H W T = 0 ; (31) PW=0,\qquad W^{T}=-W,\qquad WHW^{T}=0; \tag*{(31)} P W = 0 , W T = − W , W H W T = 0 ; (31)
the last because the order-2 d 2d 2 d term of the upper-block identity has no other possible cancellation. So the rank of W W W is even and its image is H H H -isotropic. For ν = 8 \nu=8 ν = 8 the budget reads D 2 ∣ det P D^2\mid\det P D 2 ∣ det P and deg D ≤ 4 \deg D\leq4 deg D ≤ 4 .
Real pole types and the intrinsic denominator
We keep the notation of § 10 : real nonzero Jacobi operators have rank ν ∈ { 7 , 8 } \nu\in\{7,8\} ν ∈ { 7 , 8 } and k = 16 − ν k=16-\nu k = 16 − ν . The first lemma extends the dimension counts of [24 Lemma 3.8] to both ranks. We use Rakić duality [27 §§1–2] ; it applies to the full kernel after an eigenvalue is shifted to zero, including the radial direction: for a = c X + b a=cX+b a = c X + b with b ⊥ X b\perp X b ⊥ X in the μ ∥ X ∥ 2 \mu\lVert X\rVert^2 μ ∥ X ∥ 2 -eigenspace of the unshifted J X \mathcal{J}_X J X , duality and the curvature symmetries give for the unshifted tensor J a X = − c μ ∥ X ∥ 2 b + μ ∥ b ∥ 2 X = μ ( ∥ a ∥ 2 X − ⟨ a , X ⟩ a ) \mathcal{J}_aX=-c\mu\lVert X\rVert^2b+\mu\lVert b\rVert^2X=\mu(\lVert a\rVert^2X-\langle a,X\rangle a) J a X = − c μ ∥ X ∥ 2 b + μ ∥ b ∥ 2 X = μ (∥ a ∥ 2 X − ⟨ a , X ⟩ a ) , so J a X = 0 \mathcal{J}_aX=0 J a X = 0 after subtracting μ Π ( a ) \mu\Pi(a) μ Π ( a ) . Exchanging the vectors gives the converse, so a ∈ ker J X a\in\ker\mathcal{J}_X a ∈ ker J X if and only if X ∈ ker J a X\in\ker\mathcal{J}_a X ∈ ker J a .
Generic triples and real pole types
Lemma 11.1. There is an open dense set of real orthonormal triples ( X , Y , Z ) (X,Y,Z) ( X , Y , Z ) with
dim ( Im J X ∩ Im J s Y + t Z ) ≤ 1 ( ( s , t ) ≠ ( 0 , 0 ) ) , \dim(\operatorname{Im}\mathcal{J}_X\cap\operatorname{Im}\mathcal{J}_{sY+tZ})\leq1\qquad((s,t)\ne(0,0)), dim ( Im J X ∩ Im J s Y + tZ ) ≤ 1 (( s , t ) = ( 0 , 0 )) , on which, in addition, the three pairwise Jacobi-image intersections are zero and the three Jacobi kernels have zero common intersection.
Proof. Fix an orthonormal pair ( Y , Z ) (Y,Z) ( Y , Z ) and let w w w run over its unit circle. The triples ( w , a , X ) (w,a,X) ( w , a , X ) with ∥ a ∥ = ∥ X ∥ = 1 \lVert a\rVert=\lVert X\rVert=1 ∥ a ∥ = ∥ X ∥ = 1 , a ∈ ker J w a\in\ker\mathcal{J}_w a ∈ ker J w and X ∈ ker J a X\in\ker\mathcal{J}_a X ∈ ker J a form a compact smooth iterated sphere bundle (the kernel bundles have constant rank) of dimension 1 + 2 ( k − 1 ) = 2 k − 1 1+2(k-1)=2k-1 1 + 2 ( k − 1 ) = 2 k − 1 . By Sard’s theorem and compactness the regular values of its projection to X ∈ S 15 X\in S^{15} X ∈ S 15 form an open dense set (values outside the image are regular, with empty fibre). By duality the fibre at X X X is the set of pairs ( w , a ) (w,a) ( w , a ) with a ∈ ker J X ∩ ker J w a\in\ker\mathcal{J}_X\cap\ker\mathcal{J}_w a ∈ ker J X ∩ ker J w , of dimension 2 k − 16 = 16 − 2 ν 2k-16=16-2\nu 2 k − 16 = 16 − 2 ν when nonempty and regular. For fixed w w w the unit sphere of an intersection of dimension d > 0 d>0 d > 0 embeds in it, so d − 1 ≤ 16 − 2 ν d-1\leq16-2\nu d − 1 ≤ 16 − 2 ν and dim ( Im J X ∩ Im J w ) = 2 ν − 16 + d ≤ 1 \dim(\operatorname{Im}\mathcal{J}_X\cap\operatorname{Im}\mathcal{J}_w)=2\nu-16+d\leq1 dim ( Im J X ∩ Im J w ) = 2 ν − 16 + d ≤ 1 (if d = 0 d=0 d = 0 , including an empty fibre, then ν = 8 \nu=8 ν = 8 , since for ν = 7 \nu=7 ν = 7 the two nine-dimensional kernels meet nontrivially, and this dimension is 0 0 0 ). Failure of this bound is a closed condition (a matrix-rank condition at some point of a compact circle), so the good set is open, and, by the regular-value argument, dense for each fixed ( Y , Z ) (Y,Z) ( Y , Z ) , hence on the independent-triple domain.
Orthogonalization preserves the bound: by duality, a ∈ ⋂ i ker J v i a\in\bigcap_i\ker\mathcal{J}_{v_i} a ∈ ⋂ i ker J v i if and only if span { v i } ⊂ ker J a \operatorname{span}\{v_i\}\subset\ker\mathcal{J}_a span { v i } ⊂ ker J a , so common kernels depend only on spans; projecting the plane of Y , Z Y,Z Y , Z along X X X to X ⊥ X^\perp X ⊥ keeps each span ( X , w ) \operatorname{span}(X,w) span ( X , w ) , so the common kernels, and therefore the image-intersection dimensions, agree; the Gram–Schmidt map is a surjective submersion onto orthonormal triples and sends the open dense
good set to an open dense set of orthonormal triples. The generic-pair theorem [23 Section 4, Lemma 2(1)] gives the three pairwise conditions (an intersection of three open dense sets). Finally, triples with a common unit kernel vector form the image of a compact bundle with fibre V 3 ( R k ) V_3(\mathbb{R}^k) V 3 ( R k ) over that vector, of dimension 15 + 3 k − 6 ∈ { 33 , 36 } 15+3k-6 \in\{33,36\} 15 + 3 k − 6 ∈ { 33 , 36 } , less than dim V 3 ( R 16 ) = 42 \dim V_3(\mathbb{R}^{16})=42 dim V 3 ( R 16 ) = 42 ; this closed image has measure zero.
Lemma 11.2. Let X X X be a regular anchor and ( X , Y , Z ) (X,Y,Z) ( X , Y , Z ) a triple as in Lemma 11.1 such that span ( Y , Z ) \operatorname{span}(Y,Z) span ( Y , Z ) is not contained in the polar hypersurface of F F F . Then F ( s Y + t Z ) F(sY+tZ) F ( s Y + tZ ) has no pole at any nonzero real direction.
Proof. With Ω = ( F − F 0 ) Λ \Omega=(F-F_0)\Lambda Ω = ( F − F 0 ) Λ as in Section 10, the leading Laurent coefficient W W W at a real pole is a nonzero real skew matrix, of rank at least two, and the leading coefficient of the coherent factor M X F + K X P M_XF+K_XP M X F + K X P is M X W Λ − 1 M_XW\Lambda^{-1} M X W Λ − 1 , with image of that rank in E X E_X E X . Away from the pole the columns of the factor lie in Im J s Y + t Z \operatorname{Im}J_{sY+tZ} Im J s Y + tZ , real spaces of constant dimension that vary continuously through the pole; rescaling by a power of a local parameter and passing to the limit puts the leading coefficient in the limiting Jacobi image. Its intersection with E X E_X E X then has dimension at least two, a contradiction.
Proposition 11.3. At a generic regular real anchor, let D X D_X D X be the reduced common denominator of F F F on X ⊥ X^\perp X ⊥ . If ν = 7 \nu=7 ν = 7 , it is constant or a single positive semidefinite quadratic, irreducible over R \mathbb{R} R . If ν = 8 \nu=8 ν = 8 , up to nonzero real constants it is one of
1 , Q , Q 2 , Q 1 Q 2 , f , 1,\qquad Q,\qquad Q^2,\qquad Q_1Q_2,\qquad f, 1 , Q , Q 2 , Q 1 Q 2 , f , where the Q i Q_i Q i are distinct real irreducible positive semidefinite quadrics of rank at least two and f f f is real irreducible of degree four. A rank-two quadric may split into conjugate linear factors over C \mathbb{C} C , and a real irreducible quartic into two conjugate quadrics.
Proof. The complement of the good set of Lemma 11.1 is a closed semialgebraic subset of V 3 ( R 16 ) V_3(\mathbb{R}^{16}) V 3 ( R 16 ) of dimension at most 41, with fibre over X X X in V 2 ( X ⊥ ) V_2(X^\perp) V 2 ( X ⊥ ) , of dimension 27; the anchors with a 27-dimensional fibre form a set of dimension at most 14. Outside its closure the good planes are open dense in X ⊥ X^\perp X ⊥ ; we call an anchor generic if it lies outside this set and outside the proper algebraic set of anchors excluded in the proof of Proposition 11.4, where specialization of the intrinsic denominator could create a common factor. That set is defined from the intrinsic denominator alone; the present proof does not use the second exclusion, which is made here so that both propositions use one notion of genericity. A generic real two-plane of X ⊥ X^\perp X ⊥ preserves the degree of the homogeneous D X D_X D X and its pole orders: write all entries over the least common denominator; on each irreducible component some numerator is not divisible by its defining factor, the common zero locus of these numerators on the component has complex codimension at least two, and a generic projective line avoids it and meets each reduced component transversally at smooth points (likewise for the entries of highest order at a repeated component).
A real homogeneous factor of odd degree vanishes somewhere on every real projective line on which it is not identically zero, which would give a real pole on a generic line, contrary to Lemma 11.2; so every irreducible real factor has even degree. A real indefinite quadratic has real zeros on an open set of two-planes, among which a generic one can be chosen, so it is excluded too: an irreducible real quadratic factor is semidefinite, of rank at least two, and we take it positive. Lemma 10.4 gives 2 deg D X ≤ ν 2\deg D_X\leq\nu 2 deg D X ≤ ν , leaving degree 0 or 2 for ν = 7 \nu=7 ν = 7 and 0, 2, or 4 for ν = 8 \nu=8 ν = 8 , with the listed factorizations. A polynomial irreducible over R \mathbb{R} R stays irreducible over C \mathbb{C} C or splits into two conjugate factors of equal degree.
A frame-independent polar divisor
Proposition 11.4. On the open set of pairs with nonzero norms and Im J X ∩ Im J Y = 0 \operatorname{Im} J_X \cap\operatorname{Im} J_Y = 0 Im J X ∩ Im J Y = 0 there is a unique rational matrix T ( X , Y ) T(X,Y) T ( X , Y ) , regular there, with
Im T ⊂ Im J Y , Im T T ⊂ Im J X , T + T T = 2 B ( X , Y ) . \operatorname{Im} T \subset\operatorname{Im} J_Y,\qquad\operatorname{Im} T^{T} \subset\operatorname{Im} J_X,\qquad T + T^{T} = 2B(X,Y). Im T ⊂ Im J Y , Im T T ⊂ Im J X , T + T T = 2 B ( X , Y ) . Its reduced common denominator D ( X , Y ) D(X,Y) D ( X , Y ) can be chosen real and bihomogeneous of bidegree ( d , d ) (d,d) ( d , d ) , with d ∈ { 0 , 2 } d \in\{0,2\} d ∈ { 0 , 2 } for ν = 7 \nu= 7 ν = 7 and d ∈ { 0 , 2 , 4 } d \in\{0,2,4\} d ∈ { 0 , 2 , 4 } for ν = 8 \nu= 8 ν = 8 ; it has no nonconstant factor depending on one slot only, and D ( Y , X ) = D ( X , Y ) = D ( X , Y + a X ) D(Y,X) = D(X,Y) = D(X,Y+aX) D ( Y , X ) = D ( X , Y ) = D ( X , Y + a X ) . For a generic anchor its fibre is D X D_X D X up to a nonzero constant. If d = 2 d = 2 d = 2 , one global choice of sign makes D ( X , Y ) ≥ 0 D(X,Y) \ge0 D ( X , Y ) ≥ 0 on real pairs, and D ( X , Y ) = ⟨ A X Y , Y ⟩ D(X,Y) = \langle A_XY,Y\rangle D ( X , Y ) = ⟨ A X Y , Y ⟩ is the sectional form of an algebraic curvature tensor.
Proof. The coherent two-plane factors give T = M Y Λ M X T T = M_Y\Lambda M_X^T T = M Y Λ M X T . The difference of two solutions is skew with image in both Jacobi images, which meet only in 0, so it is zero. Writing the image conditions with the rational spectral projections (T = π Y T π X T = \pi_Y T\pi_X T = π Y T π X ) gives a uniquely solvable linear system over R ( X , Y ) \mathbb{R}(X,Y) R ( X , Y ) , so T T T is rational, and uniqueness gives T ( Y , X ) = T ( X , Y ) T T(Y,X) = T(X,Y)^T T ( Y , X ) = T ( X , Y ) T , T ( t X , s Y ) = t s T ( X , Y ) T(tX,sY) = tsT(X,Y) T ( tX , s Y ) = t s T ( X , Y ) and T ( X , Y + a X ) = T ( X , Y ) + a J X T(X,Y+aX) = T(X,Y) + aJ_X T ( X , Y + a X ) = T ( X , Y ) + a J X (first from the coherent factor, then as a rational identity), and by exchange the other shear.
Take the least common multiple of the reduced denominators of the entries of T T T ; separate homogeneity gives a bihomogeneous generator, and exchange makes the bidegrees equal. Each shear is a polynomial automorphism adding only a polynomial matrix to T T T , so it preserves the polar divisor with its orders and changes D D D by a constant factor; the part of D ( X , Y + a X ) D(X,Y+aX) D ( X , Y + a X ) of top degree in Y Y Y is D ( X , Y ) D(X,Y) D ( X , Y ) , so this factor is 1. A one-slot factor would have infinitely many distinct irreducible transforms under the other shear, which is impossible for the finitely many polar components.
For generic unit real X X X , M Y = T ( X , Y ) M X Λ − 1 M_Y = T(X,Y)M_X\Lambda^{-1} M Y = T ( X , Y ) M X Λ − 1 and T = M Y Λ M X T T = M_Y\Lambda M_X^T T = M Y Λ M X T have the same poles, those of F F F (the lower block is polynomial). For each irreducible denominator factor g g g choose a numerator N N N not divisible by g g g ; V ( g , N ) V(g,N) V ( g , N ) has codimension at least two in the pair space, so outside a proper algebraic set of anchors its fibres have codimension at least two and contain no component of the specialized denominator. So generic specialization creates no common factor, the shear removes the radial coordinate of Y Y Y , the fibre degree is d d d , and Proposition 11.3 gives its values.
The elementary shears generate S L 2 \mathrm{SL}_2 SL 2 on the two slots, so D ( Y , − X ) = D ( X , Y ) D(Y,-X) = D(X,Y) D ( Y , − X ) = D ( X , Y ) , and the even degree gives symmetry. For d = 2 d = 2 d = 2 , D ( X , a X ) = D ( X , 0 ) = 0 D(X,aX) = D(X,0) = 0 D ( X , a X ) = D ( X , 0 ) = 0 ; writing D ( X , Y ) = ⟨ A X Y , Y ⟩ D(X,Y) = \langle A_XY,Y\rangle D ( X , Y ) = ⟨ A X Y , Y ⟩ , shear invariance gives A X X = 0 A_XX = 0 A X X = 0 , and (2.2) gives the curvature tensor. For generic real X X X the form D ( X , ⋅ ) D(X,\cdot) D ( X , ⋅ ) is semidefinite by Proposition 11.3 , hence for all X X X by continuity. Orient a nonzero fibre positively at X 0 X_0 X 0 ; for Y Y Y outside its kernel, D ( Y , X 0 ) = D ( X 0 , Y ) > 0 D(Y,X_0) = D(X_0,Y) > 0 D ( Y , X 0 ) = D ( X 0 , Y ) > 0 , so D ( Y , ⋅ ) D(Y,\cdot) D ( Y , ⋅ ) is positive semidefinite, and density and continuity give the global sign.
The intrinsic divisor is used only at generic anchors, where it specializes to the polar divisor of F F F .
Auxiliary lemmas on the transpose pencil
This section collects arguments used in several of the later case exclusions. Throughout, ν = 8 \nu= 8 ν = 8 , the anchor is regular, and the transpose pencil of Section 10 is written B ( a ) B(a) B ( a ) , with Gram matrix G ( a ) = B ( a ) B ( a ) T G(a) = B(a)B(a)^T G ( a ) = B ( a ) B ( a ) T . By Lemma 10.2 and (10.8) it satisfies the following conditions.
Setting (T). V = R 16 = E ⊕ E ⊥ V = \mathbb{R}^{16} = E \oplus E^\perp V = R 16 = E ⊕ E ⊥ with dim E = 8 \dim E = 8 dim E = 8 ; the map B : R 8 → Hom ( R 8 , V ) B:\mathbb{R}^8 \to\operatorname{Hom}(\mathbb{R}^8,V) B : R 8 → Hom ( R 8 , V ) is real linear and rank B ( a ) = 8 B(a) = 8 B ( a ) = 8 for every real a ≠ 0 a \ne0 a = 0 ; there are a linear isomorphism L : R 8 → E L:\mathbb{R}^8 \to E L : R 8 → E , an isometry K : R 8 → E ⊥ K:\mathbb{R}^8 \to E^\perp K : R 8 → E ⊥ and a vector g 0 ≠ 0 g_0 \ne0 g 0 = 0 such that
B ( a ) g 0 = − L a , K T B ( a ) is skew-symmetric ( a ∈ R 8 ) . (32) B(a)g_0 = -La,\qquad K^T B(a)\text{ is skew-symmetric}\qquad(a \in\mathbb{R}^8). \tag*{(32)} B ( a ) g 0 = − L a , K T B ( a ) is skew-symmetric ( a ∈ R 8 ) . (32)
At a regular anchor X X X one takes E = E X E=E_X E = E X , L = M X L=M_X L = M X , K = K X K=K_X K = K X and g 0 = g X g_0=g_X g 0 = g X . A real linear change of the coefficient variable a a a preserves (T), with L L L replaced by its composite with the change; L L L need not stay an isometry. On coefficient subspaces we write ∣ x ∣ 2 |x|^2 ∣ x ∣ 2 for a Euclidean quadratic form and also for its complex-bilinear extension to complex vectors.
Positive semidefinite pencils
Lemma 12.1. Let D 0 , D 1 D_0,D_1 D 0 , D 1 be real symmetric matrices and let I ⊂ R I \subset\mathbb{R} I ⊂ R be an interval with nonempty interior such that D ( t ) = D 1 + t D 0 ⪰ 0 D(t)=D_1+tD_0 \succeq0 D ( t ) = D 1 + t D 0 ⪰ 0 for all t ∈ I t \in I t ∈ I . Then ker D ( t ) = ker D 0 ∩ ker D 1 \ker D(t)=\ker D_0 \cap\ker D_1 ker D ( t ) = ker D 0 ∩ ker D 1 for every interior point t t t of I I I . Hence D ( t ) D(t) D ( t ) has one image F = Im D 0 + Im D 1 F=\operatorname{Im}D_0+\operatorname{Im}D_1 F = Im D 0 + Im D 1 for all interior t t t . If rank D ( t ) = rank D ( t ′ ) \operatorname{rank}D(t)=\operatorname{rank}D(t') rank D ( t ) = rank D ( t ′ ) for an endpoint t t t and an interior point t ′ t' t ′ , then ker D ( t ) = ker D 0 ∩ ker D 1 \ker D(t)=\ker D_0 \cap\ker D_1 ker D ( t ) = ker D 0 ∩ ker D 1 as well.
Proof. The kernel of a sum of two positive semidefinite matrices is the intersection of their kernels. An interior point t t t is a strict convex combination of two points t 1 < t < t 2 t_1<t<t_2 t 1 < t < t 2 of I I I , and D D D is affine in t t t , so ker D ( t ) = ker D ( t 1 ) ∩ ker D ( t 2 ) \ker D(t)=\ker D(t_1)\cap\ker D(t_2) ker D ( t ) = ker D ( t 1 ) ∩ ker D ( t 2 ) . A vector killed by D ( t 1 ) D(t_1) D ( t 1 ) and D ( t 2 ) D(t_2) D ( t 2 ) with t 1 ≠ t 2 t_1\ne t_2 t 1 = t 2 is killed by D 0 D_0 D 0 and D 1 D_1 D 1 ; the converse inclusion is clear. For symmetric matrices ( ker D 0 ∩ ker D 1 ) ⊥ = Im D 0 + Im D 1 (\ker D_0\cap\ker D_1)^\perp=\operatorname{Im}D_0+\operatorname{Im}D_1 ( ker D 0 ∩ ker D 1 ) ⊥ = Im D 0 + Im D 1 . At an endpoint ker D ( t ) ⊇ ker D 0 ∩ ker D 1 \ker D(t)\supseteq\ker D_0\cap\ker D_1 ker D ( t ) ⊇ ker D 0 ∩ ker D 1 , and equal ranks give equality.
Lemma 12.2. Let C ( a ) C(a) C ( a ) , a ∈ R n a \in\mathbb{R}^n a ∈ R n , be a real symmetric matrix of any size whose entries are quadratic forms in a a a , positive semidefinite of one fixed positive rank for every a ≠ 0 a\ne0 a = 0 . Suppose that its scalar quadratic entries span a real vector space of dimension at most three and n ≥ 3 n\ge3 n ≥ 3 , or of dimension at most two and n ≥ 2 n\ge2 n ≥ 2 . Then Im C ( a ) \operatorname{Im}C(a) Im C ( a ) is independent of a ≠ 0 a\ne0 a = 0 .
Proof. The form q ( a ) = tr C ( a ) q(a)=\operatorname{tr}C(a) q ( a ) = tr C ( a ) is positive definite and lies in the scalar-entry span; choose quadratic forms p , r p,r p , r , possibly zero, such that q , p , r q,p,r q , p , r span a space containing that span, and write C = q D 0 + p D 1 + r D 2 C=qD_0+pD_1+rD_2 C = q D 0 + p D 1 + r D 2 with constant symmetric D i D_i D i . After a real change of coordinates q ( a ) = ∥ a ∥ 2 q(a)=\lVert a\rVert^2 q ( a ) = ∥ a ∥ 2 . If the span has dimension at most three and n ≥ 3 n\ge3 n ≥ 3 , the joint range of ( p , r ) (p,r) ( p , r ) on the unit sphere is convex by Brickman’s theorem [20 Theorem 1.1] , so { C ( a ) : q ( a ) = 1 } \{C(a):q(a)=1\} { C ( a ) : q ( a ) = 1 } is convex. If the span has dimension at most two (so r = 0 r=0 r = 0 ) and n ≥ 2 n\ge2 n ≥ 2 , the range of p p p on the connected unit sphere is an interval, with the same consequence. The midpoint of two members of this convex family is a member of the same rank, so the two members have the same kernel, as in the proof of Lemma 12.1 . Symmetry gives a common image, and homogeneity extends it to all a ≠ 0 a\ne0 a = 0 .
Remark 12.3. If C = G C=G C = G for a pencil in Setting (T), the fixed image in Lemma 12.2 is E E E , by Lemma 12.5 (1) with U = R 8 U=\mathbb{R}^8 U = R 8 ; over the reals Im G ( a ) = Im B ( a ) \operatorname{Im}G(a)=\operatorname{Im}B(a) Im G ( a ) = Im B ( a ) .
Fixed images and the Jacobi image
Lemma 12.4. The Pfaffian A 12 A 34 − A 13 A 24 + A 14 A 23 A_{12}A_{34}-A_{13}A_{24}+A_{14}A_{23} A 12 A 34 − A 13 A 24 + A 14 A 23 of real skew 4 × 4 4\times4 4 × 4 matrices is a quadratic form of signature ( 3 , 3 ) (3,3) ( 3 , 3 ) . Hence no four-dimensional real space of skew 4 × 4 4\times4 4 × 4 matrices consists, apart from 0 0 0 , of invertible matrices.
Proof. Each of the three products is a hyperbolic plane in its two variables, so the form has signature ( 3 , 3 ) (3,3) ( 3 , 3 ) . A four-dimensional space whose nonzero members are invertible carries a nowhere-vanishing restriction of the Pfaffian, which would then be definite on a four-dimensional subspace; a form of signature ( 3 , 3 ) (3,3) ( 3 , 3 ) has none.
Lemma 12.5. Let B B B be as in Setting (T) and let U ⊂ R 8 U\subset\mathbb{R}^8 U ⊂ R 8 be a real subspace.
If Im B ( a ) = V 1 \operatorname{Im}B(a)=V_1 Im B ( a ) = V 1 for all a ∈ U ∖ { 0 } a\in U\setminus\{0\} a ∈ U ∖ { 0 } , with V 1 V_1 V 1 fixed, and dim U ≥ 4 \dim U\ge4 dim U ≥ 4 , then V 1 = E V_1=E V 1 = E .
If Im B ( a ) ⊂ V 1 \operatorname{Im} B(a) \subset V_1 Im B ( a ) ⊂ V 1 for all a ∈ U ∖ { 0 } a \in U \setminus\{0\} a ∈ U ∖ { 0 } , with V 1 V_1 V 1 fixed, dim V 1 ≤ 9 \dim V_1 \le9 dim V 1 ≤ 9 and dim U ≥ 7 \dim U \ge7 dim U ≥ 7 , then Im B ( a ) = E \operatorname{Im} B(a) = E Im B ( a ) = E for every a ∈ U ∖ { 0 } a \in U \setminus\{0\} a ∈ U ∖ { 0 } .
Proof. In both parts the radial identity in (32) gives L ( U ) ⊂ V 1 ∩ E L(U) \subset V_1 \cap E L ( U ) ⊂ V 1 ∩ E . Put W = K T V 1 W = K^{T}V_1 W = K T V 1 ; then dim W = dim V 1 − dim ( V 1 ∩ E ) ≤ dim V 1 − dim U \dim W = \dim V_1 - \dim(V_1 \cap E) \le\dim V_1 - \dim U dim W = dim V 1 − dim ( V 1 ∩ E ) ≤ dim V 1 − dim U , and the skew map K T B ( a ) K^{T}B(a) K T B ( a ) has image in W W W for a ∈ U a \in U a ∈ U .
(1) Here dim W ≤ 4 \dim W \le4 dim W ≤ 4 , and K T B ( a ) K^{T}B(a) K T B ( a ) has image exactly W W W for every nonzero a ∈ U a \in U a ∈ U . A skew map with image W W W vanishes on W ⊥ W^{\perp} W ⊥ and is invertible on W W W , so r = dim W ∈ { 0 , 2 , 4 } r = \dim W \in\{0,2,4\} r = dim W ∈ { 0 , 2 , 4 } . If r = 2 r = 2 r = 2 , the linear map a ↦ K T B ( a ) a \mapsto K^{T}B(a) a ↦ K T B ( a ) sends U U U into the one-dimensional space of skew maps supported on W W W , so it has a nonzero kernel, contradicting rank two. If r = 4 r = 4 r = 4 , then dim U = 4 \dim U = 4 dim U = 4 and a ↦ K T B ( a ) ∣ W a \mapsto K^{T}B(a)|_W a ↦ K T B ( a ) ∣ W embeds U U U as a four-dimensional space of invertible skew maps of W W W , which Lemma 12.4 excludes. Hence r = 0 r = 0 r = 0 , so V 1 ⊂ E V_1 \subset E V 1 ⊂ E , and dim V 1 = 8 \dim V_1 = 8 dim V 1 = 8 gives V 1 = E V_1 = E V 1 = E .
(2) Here dim W ≤ 2 \dim W \le2 dim W ≤ 2 , so every K T B ( a ) K^{T}B(a) K T B ( a ) with a ∈ U a \in U a ∈ U lies in a space of skew maps of dimension at most one. Since dim U ≥ 2 \dim U \ge2 dim U ≥ 2 , some a 0 ≠ 0 a_0 \ne0 a 0 = 0 in U U U has K T B ( a 0 ) = 0 K^{T}B(a_0) = 0 K T B ( a 0 ) = 0 ; then Im B ( a 0 ) = E \operatorname{Im} B(a_0) = E Im B ( a 0 ) = E by rank eight, so E ⊂ V 1 E \subset V_1 E ⊂ V 1 and dim W ≤ 1 \dim W \le1 dim W ≤ 1 . A skew map with image of dimension at most one is zero, so K T B ( a ) = 0 K^{T}B(a) = 0 K T B ( a ) = 0 on U U U , and rank eight gives the assertion. □
Lemma 12.6. A linear space of matrices over a field, all of rank at most one, either has all images in one fixed line or has all rows in one fixed line.
Proof. Write two nonzero members as w 1 e 1 ∗ w_1e_1^{*} w 1 e 1 ∗ and w 2 e 2 ∗ w_2e_2^{*} w 2 e 2 ∗ . If w 1 , w 2 w_1,w_2 w 1 , w 2 are independent, rank at most one of the sum forces e 1 , e 2 e_1,e_2 e 1 , e 2 to be proportional; a third member w 3 e 3 ∗ w_3e_3^{*} w 3 e 3 ∗ has w 3 w_3 w 3 independent of w 1 w_1 w 1 or of w 2 w_2 w 2 , so the same argument makes e 3 e_3 e 3 proportional to e 1 e_1 e 1 or to e 2 e_2 e 2 , hence to both. If no such pair exists, the image line is common. □
Isotropic lines of a rank-two Gram matrix
Lemma 12.7. Let A ⊂ P C 7 A \subset\mathbb{P}^{7}_{\mathbb{C}} A ⊂ P C 7 be an irreducible variety in the coefficient space on which G G G has generic rank two, and let k k k be its function field (or that of its affine cone). Then K 0 = ker B K_0 = \ker B K 0 = ker B over k k k is a totally isotropic three-space for the standard bilinear form on k 8 k^8 k 8 , and B B B induces an isometry from K 0 ⊥ / K 0 K_0^{\perp}/K_0 K 0 ⊥ / K 0 onto Im G \operatorname{Im} G Im G with the pairing ⟨ G x , G x ′ ⟩ = x T G x ′ \langle Gx,Gx'\rangle= x^{T}Gx' ⟨ G x , G x ′ ⟩ = x T G x ′ . This binary pairing has determinant class − 1 -1 − 1 and is hyperbolic over k k k . If A A A is real-defined, the same holds over its real function field; then its two isotropic lines are individually defined over the reals.
Proof. Lemma 10.3 gives rank G = 2 rank B − 8 \operatorname{rank} G = 2\operatorname{rank} B - 8 rank G = 2 rank B − 8 , so rank B = 5 \operatorname{rank} B = 5 rank B = 5 and K 0 K_0 K 0 has dimension three; it is totally isotropic by the same lemma. Since Im B T = K 0 ⊥ \operatorname{Im} B^{T} = K_0^{\perp} Im B T = K 0 ⊥ , we have Im G = B ( K 0 ⊥ ) \operatorname{Im} G = B(K_0^{\perp}) Im G = B ( K 0 ⊥ ) and x T G x ′ = ( B T x ) T ( B T x ′ ) x^{T}Gx' = (B^{T}x)^{T}(B^{T}x') x T G x ′ = ( B T x ) T ( B T x ′ ) , which is the stated isometry. The standard form on k 8 k^8 k 8 is the orthogonal sum of three hyperbolic planes containing K 0 K_0 K 0 and K 0 ⊥ / K 0 K_0^{\perp}/K_0 K 0 ⊥ / K 0 ; its determinant class is 1 1 1 , so that of K 0 ⊥ / K 0 K_0^{\perp}/K_0 K 0 ⊥ / K 0 is − 1 -1 − 1 . A nondegenerate binary form of determinant class − 1 -1 − 1 is hyperbolic. For a real-defined A A A the same computation runs over the real function field. □
Lemma 12.8. Let S S S be a normal standard graded domain over C \mathbb{C} C with degree-zero part C \mathbb{C} C , and let C C C be a symmetric N × N N \times N N × N matrix of homogeneous quadratic elements of S S S with generic rank two, whose induced pairing ⟨ C x , C x ′ ⟩ = x T C x ′ \langle Cx,Cx'\rangle= x^{T}Cx' ⟨ C x , C x ′ ⟩ = x T C x ′ on its image is hyperbolic over Frac ( S ) \operatorname{Frac}(S) Frac ( S ) . Suppose that the saturated rank-one submodules of S N S^N S N along its two isotropic lines are principal. Then
C = t ( p r T + r p T ) (33) C = t(pr^{T} + rp^{T}) \tag*{(33)} C = t ( p r T + r p T ) (33)
with homogeneous t ∈ S t \in S t ∈ S and homogeneous vectors p , r p,r p , r of degrees d p , d r ≥ 0 d_p,d_r \ge0 d p , d r ≥ 0 , and deg t + d p + d r = 2 \deg t + d_p + d_r = 2 deg t + d p + d r = 2 . If moreover S = S R ⊗ R C S = S_{\mathbb{R}} \otimes_{\mathbb{R}} \mathbb{C} S = S R ⊗ R C for a geometrically integral normal standard graded real domain S R S_{\mathbb{R}} S R , the entries of C C C lie in S R S_{\mathbb{R}} S R , and the pairing is hyperbolic over Frac ( S R ) \operatorname{Frac}(S_{\mathbb{R}}) Frac ( S R ) , then p , r , t p,r,t p , r , t may be taken real, and over S R S_{\mathbb{R}} S R
C = u u T − v v T or C = c b T + b c T , (34) C = uu^{T} - vv^{T} \qquad\text{or} \qquad C = cb^{T} + bc^{T}, \tag*{(34)} C = u u T − v v T or C = c b T + b c T , (34)
with u , v u,v u , v homogeneous linear, or c c c a nonzero constant vector and b b b homogeneous quadratic.
Proof. The scaling σ τ : a ↦ τ a \sigma_{\tau}:a\mapsto\tau a σ τ : a ↦ τ a , τ ∈ C ∗ \tau\in\mathbb{C}^{*} τ ∈ C ∗ , multiplies C C C by τ 2 \tau^{2} τ 2 and so permutes the two isotropic lines. The stabilizer of a line has index at most two, and the divisible group C ∗ \mathbb{C}^{*} C ∗ has no subgroup of index two, so each saturated module is σ τ \sigma_{\tau} σ τ -stable and therefore graded. A principal graded rank-one module has a homogeneous generator (graded Nakayama at S > 0 S_{>0} S > 0 ); call the two generators p , r p,r p , r . Over the fraction field, Im C \operatorname{Im} C Im C is spanned by p , r p,r p , r , so C = [ p r ] M [ p r ] T C=[p\ r]M[p\ r]^{T} C = [ p r ] M [ p r ] T with M M M symmetric and invertible; in the basis ( p , r ) (p,r) ( p , r ) the induced pairing has Gram matrix M − 1 M^{-1} M − 1 , and isotropy of p p p and r r r makes M − 1 M^{-1} M − 1 , hence M M M , antidiagonal: C = t ( p r T + r p T ) C=t(pr^{T}+rp^{T}) C = t ( p r T + r p T ) with 0 ≠ t ∈ Frac ( S ) 0\ne t\in\operatorname{Frac}(S) 0 = t ∈ Frac ( S ) . At each height-one prime both generators are primitive, so their reductions are nonzero, and p r T + r p T pr^{T}+rp^{T} p r T + r p T has a nonzero reduction in characteristic zero (if u ≠ 0 u\ne0 u = 0 and u v T + v u T = 0 uv^{T}+vu^{T}=0 u v T + v u T = 0 , a nonzero coordinate of u u u , its diagonal entry and its row force v = 0 v=0 v = 0 ). Hence some entry of p r T + r p T pr^{T}+rp^{T} p r T + r p T is a local unit, t t t has no negative height-one valuation, and normality gives t ∈ S t\in S t ∈ S . Homogeneity gives the degree identity.
In the real case the two lines are individually real-defined, so conjugation preserves each saturated module and changes its generator by a unit of S S S , a nonzero constant; conjugating twice shows it has modulus one, and a constant phase makes the generator real. Then t t t is real. If d p , d r > 0 d_p,d_r>0 d p , d r > 0 , both equal one and t t t is a real constant; with u = ( t p + r ) / 2 u=(tp+r)/\sqrt{2} u = ( tp + r ) / 2 and v = ( t p − r ) / 2 v=(tp-r)/\sqrt{2} v = ( tp − r ) / 2 one has t ( p r T + r p T ) = u u T − v v T t(pr^{T}+rp^{T})=uu^{T}-vv^{T} t ( p r T + r p T ) = u u T − v v T . Otherwise one generator is a constant c c c and the other, multiplied by t t t , is a quadratic b b b .
Complex arcs through a quadric stratum
Lemma 12.9. Let B B B be as in Setting ( T ) (T) ( T ) . Split the coefficient space as R 8 = W ⊕ Z \mathbb{R}^{8}=W\oplus Z R 8 = W ⊕ Z , a = ( x , z ) a=(x,z) a = ( x , z ) , with w = dim W ≥ 4 w=\dim W\ge4 w = dim W ≥ 4 , put q ( x , z ) = ∣ x ∣ 2 + ∣ z ∣ 2 q(x,z)=|x|^{2}+|z|^{2} q ( x , z ) = ∣ x ∣ 2 + ∣ z ∣ 2 for Euclidean norms on W W W and Z Z Z , and let ρ \rho ρ be a real quadratic form on Z Z Z . Extend ∣ ⋅ ∣ 2 |\cdot|^{2} ∣ ⋅ ∣ 2 complex bilinearly, and suppose:
(i) G G G vanishes on Σ = { ( x , 0 ) : ∣ x ∣ 2 = 0 } ⊂ W C \Sigma=\{(x,0):|x|^{2}=0\}\subset W_{\mathbb{C}} Σ = {( x , 0 ) : ∣ x ∣ 2 = 0 } ⊂ W C ;
(ii) rank G ≤ 2 \operatorname{rank}G\le2 rank G ≤ 2 at every point of V ( q , ρ ) ⊂ ( W ⊕ Z ) C V(q,\rho)\subset(W\oplus Z)_{\mathbb{C}} V ( q , ρ ) ⊂ ( W ⊕ Z ) C .
Then the following hold.
Im B ( x , 0 ) = E \operatorname{Im}B(x,0)=E Im B ( x , 0 ) = E for every real x ∈ W ∖ { 0 } x\in W\setminus\{0\} x ∈ W ∖ { 0 } , and K T B ( x , z ) = T ( z ) K^{T}B(x,z)=T(z) K T B ( x , z ) = T ( z ) for a real linear map T T T from Z Z Z to skew matrices. For a real unit x ∗ ∈ W x_{*}\in W x ∗ ∈ W the matrix S = B ( x ∗ , 0 ) S=B(x_{*},0) S = B ( x ∗ , 0 ) maps R 8 \mathbb{R}^{8} R 8 onto E E E , and Ξ ( x ) = S − 1 B ( x , 0 ) \Xi(x)=S^{-1}B(x,0) Ξ ( x ) = S − 1 B ( x , 0 ) satisfies Ξ ( x ) T Ξ ( x ) = Ξ ( x ) Ξ ( x ) T = ∣ x ∣ 2 I \Xi(x)^{T}\Xi(x)=\Xi(x)\Xi(x)^{T}=|x|^{2}I Ξ ( x ) T Ξ ( x ) = Ξ ( x ) Ξ ( x ) T = ∣ x ∣ 2 I and Ξ ( x ∗ ) = I \Xi(x_{*})=I Ξ ( x ∗ ) = I ; the operators Γ y = Ξ ( y ) \Gamma_{y}=\Xi(y) Γ y = Ξ ( y ) , y ∈ W y\in W y ∈ W , y ⊥ x ∗ y\perp x_{*} y ⊥ x ∗ , satisfy Γ y T = − Γ y \Gamma_{y}^{T}=-\Gamma_{y} Γ y T = − Γ y and Γ y Γ y ′ + Γ y ′ Γ y = − 2 ⟨ y , y ′ ⟩ I \Gamma_{y}\Gamma_{y'}+\Gamma_{y'}\Gamma_{y}=-2\langle y,y'\rangle I Γ y Γ y ′ + Γ y ′ Γ y = − 2 ⟨ y , y ′ ⟩ I .
For every nonzero complex z 0 ∈ Z C z_{0}\in Z_{\mathbb{C}} z 0 ∈ Z C with ρ ( z 0 ) = 0 \rho(z_{0})=0 ρ ( z 0 ) = 0 , either T ( z 0 ) = 0 T(z_{0})=0 T ( z 0 ) = 0 , or U = Im T ( z 0 ) T ⊂ C 8 U=\operatorname{Im}T(z_{0})^{T}\subset\mathbb{C}^{8} U = Im T ( z 0 ) T ⊂ C 8 has dimension two, is invariant under every Γ y \Gamma_{y} Γ y , and tr ( Γ y ∣ U ) = 0 \operatorname{tr}(\Gamma_{y}|_{U})=0 tr ( Γ y ∣ U ) = 0 .
If w ≥ 5 w\ge5 w ≥ 5 , dim Z ≥ 2 \dim Z\ge2 dim Z ≥ 2 and ρ \rho ρ is nondegenerate, then T = 0 T=0 T = 0 and Im B ( a ) = E \operatorname{Im}B(a)=E Im B ( a ) = E for every real a ≠ 0 a\ne0 a = 0 .
Proof. (1) The quadric ∣ x ∣ 2 = 0 |x|^{2}=0 ∣ x ∣ 2 = 0 in W C W_{\mathbb{C}} W C is irreducible, so by (i) every entry of G ( x , 0 ) G(x,0) G ( x , 0 ) is a multiple of ∣ x ∣ 2 |x|^{2} ∣ x ∣ 2 : G ( x , 0 ) = ∣ x ∣ 2 D G(x,0)=|x|^{2}D G ( x , 0 ) = ∣ x ∣ 2 D with D D D real symmetric, positive semidefinite of rank eight. All real x ≠ 0 x\ne0 x = 0 in W W W give the same image Im D \operatorname{Im}D Im D , which is E E E by Lemma 12.5(1). Thus K T B ( x , 0 ) = 0 K^{T}B(x,0)=0 K T B ( x , 0 ) = 0 and K T B ( x , z ) = K T B ( 0 , z ) = T ( z ) K^{T}B(x,z)=K^{T}B(0,z)=T(z) K T B ( x , z ) = K T B ( 0 , z ) = T ( z ) , skew by (12.1). Since S S T = G ( x ∗ , 0 ) = D SS^{T}=G(x_{*},0)=D S S T = G ( x ∗ , 0 ) = D , we get Ξ ( x ) Ξ ( x ) T = ∣ x ∣ 2 I \Xi(x)\Xi(x)^{T}=|x|^{2}I Ξ ( x ) Ξ ( x ) T = ∣ x ∣ 2 I ; for square matrices also Ξ ( x ) T Ξ ( x ) = ∣ x ∣ 2 I \Xi(x)^{T}\Xi(x)=|x|^{2}I Ξ ( x ) T Ξ ( x ) = ∣ x ∣ 2 I . Polarizing at x ∗ + y x_{*}+y x ∗ + y and y + y ′ y+y' y + y ′ gives the relations for Γ y \Gamma_{y} Γ y .
(2) Fix also a nonzero complex x 0 ∈ W C x_0 \in W_{\mathbb{C}} x 0 ∈ W C with ∣ x 0 ∣ 2 = 0 |x_0|^2 = 0 ∣ x 0 ∣ 2 = 0 . As the gradient of ∣ x ∣ 2 |x|^2 ∣ x ∣ 2 at x 0 x_0 x 0 is nonzero, there is an analytic arc x ( t ) = x 0 + O ( t 2 ) x(t) = x_0 + O(t^2) x ( t ) = x 0 + O ( t 2 ) with ∣ x ( t ) ∣ 2 = − t 2 ∣ z 0 ∣ 2 |x(t)|^2 = -t^2|z_0|^2 ∣ x ( t ) ∣ 2 = − t 2 ∣ z 0 ∣ 2 , so a ( t ) = ( x ( t ) , t z 0 ) ∈ V ( q , ρ ) a(t) = (x(t),tz_0) \in V(q,\rho) a ( t ) = ( x ( t ) , t z 0 ) ∈ V ( q , ρ ) and G ( a ( 0 ) ) = 0 G(a(0)) = 0 G ( a ( 0 )) = 0 by (i). The coefficient of t t t in G ( a ( t ) ) G(a(t)) G ( a ( t )) is
G 1 = B ( x 0 , 0 ) B ( 0 , z 0 ) T + B ( 0 , z 0 ) B ( x 0 , 0 ) T G_1 = B(x_0,0)B(0,z_0)^{\mathsf T} + B(0,z_0)B(x_0,0)^{\mathsf T} G 1 = B ( x 0 , 0 ) B ( 0 , z 0 ) T + B ( 0 , z 0 ) B ( x 0 , 0 ) T
The three-minors of G ( a ( t ) ) G(a(t)) G ( a ( t )) vanish by (ii), so their t 3 t^3 t 3 coefficients give rank G 1 ≤ 2 \operatorname{rank} G_1 \le2 rank G 1 ≤ 2 . In the splitting V = E ⊕ E ⊥ V = E \oplus E^\perp V = E ⊕ E ⊥ the lower diagonal block of G 1 G_1 G 1 is zero and its off-diagonal block is B ( x 0 , 0 ) T ( z 0 ) T B(x_0,0)T(z_0)^{\mathsf T} B ( x 0 , 0 ) T ( z 0 ) T . A symmetric block matrix whose lower diagonal block is zero has rank at least twice the rank of its off-diagonal block, hence
rank ( Ξ ( x 0 ) T ( z 0 ) T ) ≤ 1. (35) \operatorname{rank}\bigl(\Xi(x_0)T(z_0)^{\mathsf T}\bigr) \le1. \tag*{(35)} rank ( Ξ ( x 0 ) T ( z 0 ) T ) ≤ 1. (35)
Suppose T ( z 0 ) ≠ 0 T(z_0) \ne0 T ( z 0 ) = 0 . Then U = Im T ( z 0 ) T U = \operatorname{Im} T(z_0)^{\mathsf T} U = Im T ( z 0 ) T has positive even dimension r r r , and by (35) Ξ ( x 0 ) ∣ U \Xi(x_0)|_U Ξ ( x 0 ) ∣ U has rank at most one for every null x 0 x_0 x 0 . The entries of ⋀ 2 Ξ ( x ) ∣ ⋀ 2 U \bigwedge^2\Xi(x)|_{\bigwedge^2 U} ⋀ 2 Ξ ( x ) ∣ ⋀ 2 U are quadratic and vanish on the irreducible quadric ∣ x ∣ 2 = 0 |x|^2 = 0 ∣ x ∣ 2 = 0 , so they are multiples of ∣ x ∣ 2 |x|^2 ∣ x ∣ 2 ; evaluation at x ∗ x_* x ∗ gives ⋀ 2 Ξ ( x ) ∣ ⋀ 2 U = ∣ x ∣ 2 incl ⋀ 2 U \bigwedge^2\Xi(x)|_{\bigwedge^2 U} = |x|^2\operatorname{incl}_{\bigwedge^2 U} ⋀ 2 Ξ ( x ) ∣ ⋀ 2 U = ∣ x ∣ 2 incl ⋀ 2 U . Differentiating at x ∗ x_* x ∗ in a direction y ⊥ x ∗ y \perp x_* y ⊥ x ∗ gives Γ y u ∧ v + u ∧ Γ y v = 0 \Gamma_yu \wedge v + u \wedge\Gamma_yv = 0 Γ y u ∧ v + u ∧ Γ y v = 0 for u , v ∈ U u,v \in U u , v ∈ U , and the components outside U U U show Γ y U ⊂ U \Gamma_yU \subset U Γ y U ⊂ U . If r ≥ 3 r \ge3 r ≥ 3 , the induced action of End ( U ) \operatorname{End}(U) End ( U ) on ⋀ 2 U \bigwedge^2 U ⋀ 2 U is faithful (a third basis vector detects each off-diagonal entry, and diagonal entries with d i + d j = 0 d_i + d_j = 0 d i + d j = 0 for all pairs vanish), so Γ y ∣ U = 0 \Gamma_y|_U = 0 Γ y ∣ U = 0 , contrary to Γ y 2 = − ∣ y ∣ 2 I \Gamma_y^2 = -|y|^2 I Γ y 2 = − ∣ y ∣ 2 I . Hence r = 2 r = 2 r = 2 , and the identity says tr ( Γ y ∣ U ) = 0 \operatorname{tr}(\Gamma_y|_U) = 0 tr ( Γ y ∣ U ) = 0 .
(3) If T ( z 0 ) ≠ 0 T(z_0) \ne0 T ( z 0 ) = 0 , choose orthonormal y 1 , … , y 4 ∈ W y_1,\ldots,y_4 \in W y 1 , … , y 4 ∈ W orthogonal to x ∗ x_* x ∗ and put Γ i = Γ y i \Gamma_i = \Gamma_{y_i} Γ i = Γ y i . By (1) and (2), tr ( ( Γ i ∣ U ) ( Γ j ∣ U ) ) = − 2 δ i j \operatorname{tr}((\Gamma_i|_U)(\Gamma_j|_U)) = -2\delta_{ij} tr (( Γ i ∣ U ) ( Γ j ∣ U )) = − 2 δ ij , so the four restrictions are linearly independent; but they lie in the three-dimensional space of traceless endomorphisms of U U U . Hence T ( z 0 ) = 0 T(z_0) = 0 T ( z 0 ) = 0 for every complex null vector of ρ \rho ρ . These span Z C Z_{\mathbb C} Z C when ρ \rho ρ is nondegenerate and dim Z ≥ 2 \dim Z \ge2 dim Z ≥ 2 , so T = 0 T = 0 T = 0 , K T B = 0 K^{\mathsf T}B = 0 K T B = 0 , and rank eight gives the image.
The rank-seven indefinite case
This section treats the five patterns left open by Theorems C , E and F in which a reduced Jacobi eigenvalue has multiplicity eight, namely ( 8 , 4 , 3 ) (8,4,3) ( 8 , 4 , 3 ) , ( 8 , 3 , 3 , 1 ) (8,3,3,1) ( 8 , 3 , 3 , 1 ) , ( 8 , 3 , 2 , 2 ) (8,3,2,2) ( 8 , 3 , 2 , 2 ) , ( 8 , 3 , 2 , 1 , 1 ) (8,3,2,1,1) ( 8 , 3 , 2 , 1 , 1 ) and ( 8 , 3 , 1 4 ) (8,3,1^4) ( 8 , 3 , 1 4 ) , in the case where that eigenvalue is neither the smallest nor the largest.
Proposition 13.1. Suppose the nonzero Jacobi rank is seven and Λ \Lambda Λ is indefinite. At a regular anchor the rational matrix F F F of (27) has no divisorial pole. Consequently the tensor is Clifford.
Proof. Here ν = 7 \nu= 7 ν = 7 and k = 9 k = 9 k = 9 , so P P P is 9 × 7 9 \times7 9 × 7 . For 0 ≠ a ∈ C 7 0 \ne a \in\mathbb{C}^7 0 = a ∈ C 7 let s ( a ) = rank B ( a ) s(a) = \operatorname{rank} B(a) s ( a ) = rank B ( a ) (Section 10); Lemma 10.3 gives s ( a ) ≥ 5 s(a) \ge5 s ( a ) ≥ 5 , and Lemma 10.2 gives s ( a ) = 9 s(a) = 9 s ( a ) = 9 for real a ≠ 0 a \ne0 a = 0 . Suppose there is a pole, and let Q Q Q be an irreducible complex component of its hypersurface in X C ⊥ ≅ C 15 X_{\mathbb{C}}^\perp\cong\mathbb{C}^{15} X C ⊥ ≅ C 15 , of dimension fourteen (the argument is componentwise, so a quadratic that splits over C \mathbb{C} C is no exception). The leading Laurent coefficient W W W is nonzero and skew over its function field, of rank r ≥ 2 r \ge2 r ≥ 2 . The closure of the incidence
I = { ( Y , [ a ] ) : Y generic in Q , [ a ] ∈ P ( Im W ( Y ) ) } ‾ ⊂ C 15 × P 6 \mathcal{I} = \overline{\{(Y,[a]) : Y\text{ generic in }Q,\ [a] \in\mathbb{P}(\operatorname{Im}W(Y))\}} \subset\mathbb{C}^{15} \times\mathbb{P}^6 I = {( Y , [ a ]) : Y generic in Q , [ a ] ∈ P ( Im W ( Y ))} ⊂ C 15 × P 6
is irreducible of dimension 14 + r − 1 ≥ 15 14 + r - 1 \ge15 14 + r − 1 ≥ 15 ; let Z \mathcal{Z} Z be the Zariski closure of its projection to P 6 \mathbb{P}^6 P 6 and α = dim Z \alpha= \dim\mathcal{Z} α = dim Z . By W T W = 0 W^{\mathsf T}W = 0 W T W = 0 , P ( Y ) T a = 0 P(Y)^{\mathsf T}a = 0 P ( Y ) T a = 0 on I \mathcal{I} I ; for fixed [ a ] [a] [ a ] this is a linear space in Y Y Y of dimension 15 − s ( a ) 15-s(a) 15 − s ( a ) (its annihilator is Im B ( a ) \operatorname{Im} B(a) Im B ( a ) ). The fibre dimension theorem gives 15 − α ≤ dim I − α ≤ 15 − s ( a ) 15-\alpha\le\dim\mathcal{I}-\alpha\le15-s(a) 15 − α ≤ dim I − α ≤ 15 − s ( a ) , so 5 ≤ s ( a ) ≤ α 5 \le s(a) \le\alpha 5 ≤ s ( a ) ≤ α at a generic point of Z \mathcal{Z} Z .
By W H W T = 0 W H W^{\mathsf T} = 0 W H W T = 0 , Z \mathcal{Z} Z lies in the smooth irreducible five-dimensional quadric N H = { [ a ] ∈ P 6 : a T H a = 0 } \mathcal{N}_H = \{[a] \in\mathbb{P}^6 : a^{\mathsf T}Ha = 0\} N H = {[ a ] ∈ P 6 : a T H a = 0 } , H = Λ − 1 H = \Lambda^{-1} H = Λ − 1 . So α = 5 \alpha= 5 α = 5 and Z = N H \mathcal{Z} = \mathcal{N}_H Z = N H . The Zariski-closed locus { s ( a ) ≤ 5 } \{s(a) \le5\} { s ( a ) ≤ 5 } contains a dense subset of Z \mathcal{Z} Z , hence all of Z \mathcal{Z} Z ; it has no real point, since the real rank is nine. So N H \mathcal{N}_H N H has no real point, which contradicts the indefiniteness and nondegeneracy of H H H . There is therefore no pole, and Lemma 10.1 makes the tensor Clifford.
Exclusion of a simple quadratic pole of residue rank four
We use the rank-eight construction of § 10 with Λ \Lambda Λ indefinite, so H = Λ − 1 H=\Lambda^{-1} H = Λ − 1 is a nondegenerate indefinite form on R 8 \mathbb{R}^{8} R 8 . The lower pencil P ( Y ) P(Y) P ( Y ) is a generically invertible linear 8 × 8 8 \times8 8 × 8 matrix, Ω \Omega Ω is skew with P Ω P\Omega P Ω polynomial, and the leading Laurent coefficient W W W at a pole satisfies P W = 0 PW=0 P W = 0 and W H W T = 0 WHW^{T}=0 W H W T = 0 by (31) . By Proposition 11.3, a real irreducible quadratic pole factor Q Q Q is, up to sign, positive semidefinite of rank r Q ≥ 2 r_Q \ge2 r Q ≥ 2 .
Proposition 14.1. Let ν = 8 \nu=8 ν = 8 with H H H indefinite, and let X X X be a regular real anchor at which Proposition 11.3 applies. A simple pole of Ω \Omega Ω along a real irreducible positive semidefinite quadratic Q Q Q cannot have skew residue rank four over the fraction field of R [ Y ] / ( Q ) \mathbb{R}[Y]/(Q) R [ Y ] / ( Q ) .
Proof. A linear companion. By Lemma 10.4, Q 4 Q^4 Q 4 divides det P \det P det P ; both have degree eight, so det P = c Q 4 \det P=cQ^4 det P = c Q 4 with c ≠ 0 c\ne0 c = 0 . At the prime ( Q ) (Q) ( Q ) the four-dimensional residue image lies in ker P \ker P ker P , so at least four Smith exponents are positive; their sum is four (det P = c Q 4 \det P=cQ^4 det P = c Q 4 ), so they are four ones and four zeros, and every seven-minor is divisible by Q 3 Q^3 Q 3 . Hence P ♯ = adj P / ( c Q 3 ) P^\sharp=\operatorname{adj}P/(cQ^3) P ♯ = adj P / ( c Q 3 ) is a real linear matrix with P P ♯ = P ♯ P = Q I PP^\sharp=P^\sharp P=QI P P ♯ = P ♯ P = Q I , and modulo Q Q Q it has rank four and Im P ♯ = ker P = Im W \operatorname{Im}P^\sharp=\ker P=\operatorname{Im}W Im P ♯ = ker P = Im W . As W W W is skew, W H W T = 0 WHW^T=0 W H W T = 0 says that this image is H H H -isotropic, so ( P ♯ ) T H P ♯ = Q H 0 (P^\sharp)^T HP^\sharp=QH_0 ( P ♯ ) T H P ♯ = Q H 0 with H 0 H_0 H 0 constant. (These arguments take place over the real prime ( Q ) (Q) ( Q ) ; they also apply when a rank-two Q Q Q splits over C \mathbb{C} C .) Choose e e e with Q ( e ) = 1 Q(e)=1 Q ( e ) = 1 and replace P , P ♯ P,P^\sharp P , P ♯ by P ( e ) − 1 P P(e)^{-1}P P ( e ) − 1 P , P ♯ P ( e ) P^\sharp P(e) P ♯ P ( e ) ; this does not change H H H , both matrices are now I I I at e e e , and evaluating ( P ♯ ) T H P ♯ = Q H 0 (P^\sharp)^T HP^\sharp=QH_0 ( P ♯ ) T H P ♯ = Q H 0 at e e e gives H 0 = H H_0=H H 0 = H :
P P ♯ = P ♯ P = Q I , ( P ♯ ) T H P ♯ = Q H , P ( e ) = P ♯ ( e ) = I . (36) PP^\sharp=P^\sharp P=QI,\qquad(P^\sharp)^T HP^\sharp=QH,\qquad P(e)=P^\sharp(e)=I. \tag*{(36)} P P ♯ = P ♯ P = Q I , ( P ♯ ) T H P ♯ = Q H , P ( e ) = P ♯ ( e ) = I . (36)
In coordinates with Q = x 0 2 + y 1 2 + ⋯ + y r Q − 1 2 Q=x_0^2+y_1^2+\cdots+y_{r_Q-1}^2 Q = x 0 2 + y 1 2 + ⋯ + y r Q − 1 2 , e e e the first basis vector and z j z_j z j spanning the radical of Q Q Q , (14.1) gives
P = x 0 I + ∑ i = 1 r Q − 1 y i C i + ∑ j z j D j , P ♯ = x 0 I − ∑ i = 1 r Q − 1 y i C i − ∑ j z j D j . P=x_0I+\sum_{i=1}^{r_Q-1}y_iC_i+\sum_j z_jD_j,\qquad P^\sharp=x_0I-\sum_{i=1}^{r_Q-1}y_iC_i-\sum_j z_jD_j. P = x 0 I + i = 1 ∑ r Q − 1 y i C i + j ∑ z j D j , P ♯ = x 0 I − i = 1 ∑ r Q − 1 y i C i − j ∑ z j D j .
where the C i C_i C i are H H H -skew, square to − I -I − I and anticommute, and every D ∈ D = span R { D j } D\in\mathcal{D}=\operatorname{span}_{\mathbb{R}}\{D_j\} D ∈ D = span R { D j } is H H H -skew, satisfies D 2 = 0 D^2=0 D 2 = 0 , and anticommutes with every C i C_i C i and every member of D \mathcal{D} D .
A dimension bound. Let g X g_X g X be the unit vector with K X g X = X K_Xg_X=X K X g X = X (§ 10 ). The radial identity (10.8) reads g X T P ( Y ) = − Y T M X g_X^TP(Y)=-Y^TM_X g X T P ( Y ) = − Y T M X , whose values fill R 8 \mathbb{R}^{8} R 8 as Y Y Y ranges over X ⊥ X^\perp X ⊥ , since M X M_X M X is an orthonormal frame. This survives the row change, so the coefficient matrices of P P P span at least eight dimensions:
8 ≤ r Q + dim D . (37) 8\le r_Q+\dim\mathcal{D}. \tag*{(37)} 8 ≤ r Q + dim D . (37)
Signatures. Averaging a positive metric over the signed monomials in the C i C_i C i gives an invariant metric, for which the operator representing H H H commutes with the C i C_i C i ; its eigenspaces are therefore modules. A single generator makes their dimensions even, two generators make them multiples of four, and four make them multiples of eight (C 1 C 2 C 3 C_1C_2C_3 C 1 C 2 C 3 is an involution commuting with C 1 , C 2 , C 3 C_1,C_2,C_3 C 1 , C 2 , C 3 and anticommuting with C 4 C_4 C 4 ; its two eigenspaces are quaternionic under C 1 , C 2 C_1,C_2 C 1 , C 2 and exchanged by C 4 C_4 C 4 ). As H H H is indefinite on R 8 \mathbb{R}^{8} R 8 , r Q ≤ 4 r_Q\le4 r Q ≤ 4 , and up to sign either H H H has signature ( 2 , 6 ) (2,6) ( 2 , 6 ) and r Q = 2 r_Q=2 r Q = 2 , or H H H has signature ( 4 , 4 ) (4,4) ( 4 , 4 ) and r Q ∈ { 2 , 3 , 4 } r_Q\in\{2,3,4\} r Q ∈ { 2 , 3 , 4 } .
Antilinear skew maps. Identify R 8 = C 4 \mathbb{R}^{8}=\mathbb{C}^{4} R 8 = C 4 by J = C 1 J=C_1 J = C 1 . Then H H H is the real part of a Hermitian form whose signature is half the real signature; in a complex basis its matrix S H S_H S H may be chosen real diagonal or real hyperbolic. A map anticommuting with J J J is D ( v ) = Θ v ˉ D(v)=\Theta\bar{v} D ( v ) = Θ v ˉ , and H H H -skewness reads S H Θ + Θ T S H = 0 S_H\Theta+\Theta^TS_H=0 S H Θ + Θ T S H = 0 , so S H Θ S_H\Theta S H Θ is a complex alternating matrix; its rank is even, and the real rank of $D is twice the complex rank of Θ \Theta Θ , hence a multiple of four. Since D 2 = 0 D^2=0 D 2 = 0 and D D D is H H H -skew, the image of D D D is totally H H H -isotropic, so its rank is at most the real Witt index of H H H .
Signature ( 2 , 6 ) (2,6) ( 2 , 6 ) . The Witt index is 2, so every D ∈ D D\in\mathcal{D} D ∈ D is zero; D = 0 D=0 D = 0 contradicts (14.2) with r Q = 2 r_Q=2 r Q = 2 .
Signature ( 4 , 4 ) (4,4) ( 4 , 4 ) , r Q = 2 r_Q=2 r Q = 2 . Let 0 ≠ D 0 ∈ D 0\ne D_0\in\mathcal{D} 0 = D 0 ∈ D (if D = 0 \mathcal{D}=0 D = 0 , (14.2) already fails). By the paragraph on antilinear skew maps, D 0 D_0 D 0 has real rank four, so L = Im D 0 = ker D 0 L=\operatorname{Im}D_0=\ker D_0 L = Im D 0 = ker D 0 is a J J J -invariant maximal isotropic subspace. In a hyperbolic basis adapted to L L L and an isotropic complement, the invertible upper-right block of Θ 0 \Theta_0 Θ 0 is alternating, and a change of basis of L L L (with the dual change on the complement) gives
S H = [ 0 I 2 I 2 0 ] , Θ 0 = [ 0 J 2 0 0 ] , J 2 = [ 0 1 − 1 0 ] . S_H=\begin{bmatrix}0&I_2\\ I_2&0\end{bmatrix},\qquad
\Theta_0=\begin{bmatrix}0&J_2\\0&0\end{bmatrix},\qquad
J_2=\begin{bmatrix}0&1\\-1&0\end{bmatrix}. S H = [ 0 I 2 I 2 0 ] , Θ 0 = [ 0 0 J 2 0 ] , J 2 = [ 0 − 1 1 0 ] .
Every D ∈ D D\in\mathcal{D} D ∈ D preserves L L L , by D D 0 = − D 0 D DD_0=-D_0D D D 0 = − D 0 D ; skewness gives Θ = [ Θ 11 Θ 12 0 − Θ 11 T ] \Theta=\begin{bmatrix}\Theta_{11}&\Theta_{12}\\0&-\Theta_{11}^{T}\end{bmatrix} Θ = [ Θ 11 0 Θ 12 − Θ 11 T ] with Θ 12 T = − Θ 12 \Theta_{12}^{T}=-\Theta_{12} Θ 12 T = − Θ 12 , and D D 0 = − D 0 D DD_0=-D_0D D D 0 = − D 0 D gives Θ 11 J 2 = J 2 Θ ‾ 11 \Theta_{11}J_2=J_2\overline{\Theta}_{11} Θ 11 J 2 = J 2 Θ 11 , whose solutions are Θ 11 = [ a i b i c a ‾ ] \Theta_{11}=\begin{bmatrix}a&ib\\ic&\overline{a}\end{bmatrix} Θ 11 = [ a i c ib a ] with a ∈ C a\in\mathbb{C} a ∈ C , b , c ∈ R b,c\in\mathbb{R} b , c ∈ R . The block Θ 11 Θ ‾ 11 = ( ∣ a ∣ 2 + b c ) I 2 \Theta_{11}\overline{\Theta}_{11}=(|a|^2+bc)I_2 Θ 11 Θ 11 = ( ∣ a ∣ 2 + b c ) I 2 of D 2 D^2 D 2 vanishes; the form ∣ a ∣ 2 + b c |a|^2+bc ∣ a ∣ 2 + b c has signature ( 3 , 1 ) (3,1) ( 3 , 1 ) , so the image of the real linear map D ↦ Θ 11 D\mapsto\Theta_{11} D ↦ Θ 11 is a linear subspace of its null cone, hence has dimension at most one, and its kernel, where Θ 11 = 0 \Theta_{11}=0 Θ 11 = 0 and Θ 12 = f J 2 \Theta_{12}=fJ_2 Θ 12 = f J 2 with f ∈ C f\in\mathbb{C} f ∈ C , has dimension at most two. Thus r Q + dim D ≤ 5 r_Q+\dim\mathcal{D}\le5 r Q + dim D ≤ 5 , a contradiction.
Signature ( 4 , 4 ) (4,4) ( 4 , 4 ) , r Q ∈ { 3 , 4 } r_Q\in\{3,4\} r Q ∈ { 3 , 4 } . Put J ′ = C 1 C 2 J'=C_1C_2 J ′ = C 1 C 2 and T = − J ′ D T=-J'D T = − J ′ D for D ∈ D D\in\mathcal{D} D ∈ D . As J ′ J' J ′ and D D D are H H H -skew, D D D commutes with J ′ J' J ′ , and both anticommute with C 1 , C 2 C_1,C_2 C 1 , C 2 , the map T T T is H H H -self-adjoint, satisfies T 2 = 0 T^2=0 T 2 = 0 and commutes with C 1 , C 2 C_1,C_2 C 1 , C 2 , so it is quaternionic-linear. The form H H H is the real part of a quaternionic Hermitian form of signature ( 1 , 1 ) (1,1) ( 1 , 1 ) , and in a quaternionic basis with that form diagonal, T = [ a b − b ‾ d ] T=\begin{bmatrix}a&b\\-\overline{b}&d\end{bmatrix} T = [ a − b b d ] with a , d ∈ R a,d\in\mathbb{R} a , d ∈ R , b ∈ H b\in\mathbb{H} b ∈ H . Nilpotence gives zero real trace, 4 ( a + d ) = 0 4(a+d)=0 4 ( a + d ) = 0 , so d = − a d=-a d = − a and T 2 = ( a 2 − ∣ b ∣ 2 ) I 2 T^2=(a^2-|b|^2)I_2 T 2 = ( a 2 − ∣ b ∣ 2 ) I 2 . The real linear image of D \mathcal{D} D lies in the null cone of this form of signature ( 1 , 4 ) (1,4) ( 1 , 4 ) and hence has dimension at most one; the map D ↦ − J ′ D D\mapsto-J'D D ↦ − J ′ D is injective, so dim D ≤ 1 \dim\mathcal{D}\le1 dim D ≤ 1 and r Q + dim D ≤ 5 r_Q+\dim\mathcal{D}\le5 r Q + dim D ≤ 5 , a contradiction.
Exclusion of a double quadratic pole
We work in the indefinite case ν = 8 \nu=8 ν = 8 with the notation of § 10 and § 11 : at a regular anchor X X X , for Y ⊥ X Y\perp X Y ⊥ X ,
M Y = M X F ( Y ) + K X P ( Y ) , F = F 0 + Ω H , Ω T = − Ω , H = Λ − 1 . M_Y=M_XF(Y)+K_XP(Y),\qquad F=F_0+\Omega H,\qquad\Omega^T=-\Omega,\qquad H=\Lambda^{-1}. M Y = M X F ( Y ) + K X P ( Y ) , F = F 0 + Ω H , Ω T = − Ω , H = Λ − 1 .
where F 0 , P F_0,P F 0 , P are linear, P P P is generically invertible, and P Ω P\Omega P Ω is polynomial, because the lower–upper Jacobi block P Λ F T P\Lambda F^T P Λ F T is.
Proposition 15.1. There is no nonempty open set of regular anchors X X X at which the reduced common denominator of F F F on X ⊥ X^\perp X ⊥ is Q X 2 Q_X^2 Q X 2 for an irreducible real positive semidefinite quadratic Q X Q_X Q X (in particular, no branch U j U_j U j of Lemma 27.1 with nonempty interior is of this type).
A degree bound for an entire skew block
Lemma 15.2. Let P 0 P_0 P 0 be a real polynomial m × b m\times b m × b matrix with linear entries and generic column rank b b b . Let Ω 0 \Omega_0 Ω 0 be a skew rational b × b b\times b b × b matrix such that P 0 Ω 0 P_0\Omega_0 P 0 Ω 0 is polynomial. If Ω 0 \Omega_0 Ω 0 has pole order e e e at a real irreducible quadratic Q Q Q , then every maximal b b b -minor of P 0 P_0 P 0 is divisible by Q 2 e Q^{2e} Q 2 e . In particular, b ≥ 4 e b\ge4e b ≥ 4 e .
Proof. Over the discrete valuation ring at Q Q Q take a Smith form P 0 = U [ diag ( Q a 1 , … , Q a b ) ] V P_0=U\left[\operatorname{diag}(Q^{a_1},\ldots,Q^{a_b})\right]V P 0 = U [ diag ( Q a 1 , … , Q a b ) ] V with U , V U,V U , V invertible over that ring. The congruent skew matrix Ω ^ = V Ω 0 V T \widehat{\Omega}=V\Omega_0V^{T} Ω = V Ω 0 V T still has pole order e e e ; choose an off-diagonal entry Ω ^ i j \widehat{\Omega}_{ij} Ω ij of valuation − e -e − e . Polynomiality of P 0 Ω 0 P_0\Omega_0 P 0 Ω 0 makes diag ( Q a 1 , … , Q a b ) Ω ^ \operatorname{diag}(Q^{a_1},\ldots,Q^{a_b})\widehat{\Omega} diag ( Q a 1 , … , Q a b ) Ω integral; its rows i i i and j j j , which contain the entries ± Ω ^ i j \pm\widehat{\Omega}_{ij} ± Ω ij , force a i , a j ≥ e a_i,a_j\ge e a i , a j ≥ e , so every maximal minor has valuation at least 2 e 2e 2 e . Some maximal minor is nonzero, of degree at most b b b , whereas deg Q 2 e = 4 e \deg Q^{2e}=4e deg Q 2 e = 4 e .
Under the double-pole hypothesis this lemma, applied to P , Ω P,\Omega P , Ω , gives by degree eight
det P = c X Q X 4 , c X ≠ 0. (38) \det P=c_XQ_X^4,\qquad c_X\ne0. \tag*{(38)} det P = c X Q X 4 , c X = 0. (38)
Let g X g_X g X be the unit vector with K X g X = X K_Xg_X=X K X g X = X (§ 10 ). The radial row g X T P ( Y ) = − Y T M X g_X^TP(Y)=-Y^TM_X g X T P ( Y ) = − Y T M X vanishes for Y Y Y in the reduced zero eigenspace E 0 ( X ) = ker J X ∩ X ⊥ E_0(X)=\ker J_X\cap X^\perp E 0 ( X ) = ker J X ∩ X ⊥ , so ((38) ) gives Q X ( Y ) = 0 Q_X(Y)=0 Q X ( Y ) = 0 there, and semidefiniteness puts these vectors in the radical of Q X Q_X Q X . With the shear invariance of the intrinsic denominator (Proposition 11.4 ), the quadratic pole form annihilates the full kernel of J X J_X J X and has rank at most eight.
The auxiliary quadratic family
Let D ( X , Y ) D(X,Y) D ( X , Y ) be the intrinsic reduced denominator of Proposition 11.4 . Its fibre at a generic anchor of the open set in Proposition 15.1 is Q X 2 Q_X^2 Q X 2 , so by Proposition 11.4 it has bidegree ( 4 , 4 ) (4,4) ( 4 , 4 ) and no factor in one slot only, so each irreducible factor of D D D in R [ X , Y ] \mathbb{R}[X,Y] R [ X , Y ] is primitive in Y Y Y , and Gauss’s lemma identifies the factorization with the one over R ( X ) [ Y ] \mathbb{R}(X)[Y] R ( X ) [ Y ] . Distinct irreducible factors are coprime over R ( X ) \mathbb{R}(X) R ( X ) and squarefree over its algebraic closure, the characteristic being zero; after a generic linear change of the Y Y Y -coordinates their resultants and discriminants in one coordinate are nonzero polynomials, which stay nonzero off a proper algebraic set of anchors. On the nonempty Zariski-open set of anchors where these resultants and discriminants are nonzero and the fibre of D D D is the reduced denominator (Proposition 11.4 ), a set that meets that open set, each factor fibre divides Q X 2 Q_X^2 Q X 2 and is squarefree, hence is proportional to the irreducible quadratic Q X Q_X Q X , and coprimeness leaves one factor, with exponent two. Hence
D = c Q ( X , Y ) 2 (39) D=cQ(X,Y)^2 \tag*{(39)} D = c Q ( X , Y ) 2 (39)
with Q Q Q real of bidegree ( 2 , 2 ) (2,2) ( 2 , 2 ) and c ≠ 0 c\ne0 c = 0 . The two shear identities of D D D pass to Q Q Q : their squares agree in the polynomial ring with the shear parameter, the domain property gives a common sign, and parameter zero makes it positive. The shears generate ( X , Y ) ↦ ( Y , − X ) (X,Y)\mapsto(Y,-X) ( X , Y ) ↦ ( Y , − X ) , and the even bidegrees give Q ( Y , X ) = Q ( X , Y ) Q(Y,X)=Q(X,Y) Q ( Y , X ) = Q ( X , Y ) , Q ( X , Y + t X ) = Q ( X , Y ) Q(X,Y+tX)=Q(X,Y) Q ( X , Y + tX ) = Q ( X , Y ) and Q ( X , X ) = 0 Q(X,X)=0 Q ( X , X ) = 0 . Write Q ( X , Y ) = ⟨ A X Y , Y ⟩ Q(X,Y)=\langle A_XY,Y\rangle Q ( X , Y ) = ⟨ A X Y , Y ⟩ ; then A X A_X A X is a symmetric quadratic family with A X X = 0 A_XX=0 A X X = 0 .
Anchors at which Proposition 11.3 applies and the fibre of D D D is D X D_X D X are dense. At each of them c Q ( X , ⋅ ) 2 = D X cQ(X,\cdot)^2=D_X c Q ( X , ⋅ ) 2 = D X , and among the factor types of Proposition 11.3 the only square is the square of an irreducible quadratic, so Q ( X , ⋅ ) Q(X,\cdot) Q ( X , ⋅ ) is an irreducible positive semidefinite quadratic with a double pole and the argument after ((38) ) applies at X X X . Semidefiniteness persists by continuity. Fix one such anchor X 0 X_0 X 0 and the sign of Q Q Q so that Q ( X 0 , ⋅ ) ≥ 0 Q(X_0,\cdot)\ge0 Q ( X 0 , ⋅ ) ≥ 0 ; symmetry gives Q ( Y , X 0 ) = Q ( X 0 , Y ) > 0 Q(Y,X_0)=Q(X_0,Y)>0 Q ( Y , X 0 ) = Q ( X 0 , Y ) > 0 for Y Y Y outside its kernel, so A Y ≥ 0 A_Y\ge0 A Y ≥ 0 on a dense set and then everywhere. The argument after ((38) ) and continuity of the kernel projector give
ker J X ⊂ ker A X , A X ≥ 0. (40) \ker J_X\subset\ker A_X,\qquad A_X\ge0. \tag*{(40)} ker J X ⊂ ker A X , A X ≥ 0. (40)
Lemma 15.3. Let A z A_z A z be a symmetric quadratic family with Im A z ⊂ Im J z \operatorname{Im}A_z\subset\operatorname{Im}J_z Im A z ⊂ Im J z for all real z ≠ 0 z\ne0 z = 0 . For the anchor X X X and, as a rational identity, for Y ∈ X ⊥ Y\in X^\perp Y ∈ X ⊥ ,
A s X + Y = ( s M X + M Y ) C ( s M X + M Y ) T , C = M X T A X M X , A_{sX+Y}=(sM_X+M_Y)C(sM_X+M_Y)^T,\qquad C=M_X^TA_XM_X, A s X + Y = ( s M X + M Y ) C ( s M X + M Y ) T , C = M X T A X M X ,
with C C C constant, C ⪰ 0 C \succeq0 C ⪰ 0 if A ⪰ 0 A \succeq0 A ⪰ 0 , and
Ω H C − C H Ω polynomial . (41) \Omega HC-CH\Omega\quad\text{polynomial}. \tag*{(41)} Ω H C − C H Ω polynomial . (41)
Proof. For a generic partner Y Y Y the matrix M ^ = [ M X , M Y ] \widehat{M}=[M_X,M_Y] M = [ M X , M Y ] is injective, and A s X + t Y = M ^ M ( s , t ) M ^ T A_{sX+tY}=\widehat{M}\mathcal{M}(s,t)\widehat{M}^{T} A s X + t Y = M M ( s , t ) M T for a unique symmetric quadratic block matrix M \mathcal{M} M . Its values at the axes give
M = [ s 2 C + s t Θ 11 s t Θ 12 s t Θ 12 T t 2 C 1 + s t Θ 22 ] . \mathcal{M}=
\begin{bmatrix}
s^2C+st\Theta_{11} & st\Theta_{12}\\
st\Theta_{12}^{T} & t^2C_1+st\Theta_{22}
\end{bmatrix}. M = [ s 2 C + s t Θ 11 s t Θ 12 T s t Θ 12 t 2 C 1 + s t Θ 22 ] .
Since Im A s X + t Y ⊂ Im M ^ [ s I , t I ] T \operatorname{Im}A_{sX+tY}\subset\operatorname{Im}\widehat{M}[sI,tI]^T Im A s X + t Y ⊂ Im M [ s I , t I ] T , we get [ − t I , s I ] M = 0 [-tI,sI]\mathcal{M}=0 [ − t I , s I ] M = 0 for generic real s , t s,t s , t , hence identically, and comparing coefficients gives Θ 11 = Θ 22 = 0 \Theta_{11}=\Theta_{22}=0 Θ 11 = Θ 22 = 0 and Θ 12 = C 1 = C \Theta_{12}=C_1=C Θ 12 = C 1 = C . This holds on the generic pair locus and then rationally. The mixed coefficient of A s X + Y A_{sX+Y} A s X + Y makes F C + C F T FC+CF^T F C + C F T polynomial, and F = F 0 + Ω H F=F_0+\Omega H F = F 0 + Ω H gives (41) .
By (40) the lemma applies here, with rank C = rank Q X C=\operatorname{rank}Q_X C = rank Q X , since the fibre defining A X A_X A X is proportional to Q X Q_X Q X .
The polar part and total isotropy
Lemma 15.4. Let C ⪰ 0 C\succeq0 C ⪰ 0 have rank r r r , put r ′ = 8 − r r'=8-r r ′ = 8 − r , and let Ω \Omega Ω be skew rational with P Ω P\Omega P Ω and Ω H C − C H Ω \Omega HC-CH\Omega Ω H C − C H Ω polynomial. After subtracting a polynomial skew matrix from Ω \Omega Ω ,
Ω H C = C H Ω , P Ω polynomial . (42) \Omega HC=CH\Omega,\qquad P\Omega\ \text{polynomial}. \tag*{(42)} Ω H C = C H Ω , P Ω polynomial . (42) After a constant change S S S of coefficient coordinates, C ′ = S − 1 C S − T C'=S^{-1}CS^{-T} C ′ = S − 1 C S − T , H ′ = S T H S H'=S^THS H ′ = S T H S , Ω ′ = S − 1 Ω S − T \Omega'=S^{-1}\Omega S^{-T} Ω ′ = S − 1 Ω S − T , P ′ = P S P'=PS P ′ = P S (this keeps skewness, (42) , generic injectivity of column subpencils, and the lower output frame), on spaces U 1 , U 2 , U 3 , U 4 U_1,U_2,U_3,U_4 U 1 , U 2 , U 3 , U 4 of dimensions d , n 0 , n 0 , ℓ d,n_0,n_0,\ell d , n 0 , n 0 , ℓ with d = r − n 0 d=r-n_0 d = r − n 0 , ℓ = r ′ − n 0 \ell=r'-n_0 ℓ = r ′ − n 0 , n 0 ≤ min ( r , r ′ ) n_0\leq\min(r,r') n 0 ≤ min ( r , r ′ ) ,
C = diag ( I d , I n 0 , 0 n 0 , 0 ℓ ) , H = [ H 0 0 0 0 0 0 I n 0 0 0 I n 0 0 0 0 0 0 H 1 ] . (43) C=\operatorname{diag}(I_d,I_{n_0},0_{n_0},0_\ell),\qquad
H=
\begin{bmatrix}
H_0&0&0&0\\
0&0&I_{n_0}&0\\
0&I_{n_0}&0&0\\
0&0&0&H_1
\end{bmatrix}. \tag*{(43)} C = diag ( I d , I n 0 , 0 n 0 , 0 ℓ ) , H = H 0 0 0 0 0 0 I n 0 0 0 I n 0 0 0 0 0 0 H 1 . (43) Ω = [ Z 0 0 0 0 B 0 Ψ T 0 0 Ψ 0 0 0 − T 0 T 0 V 0 ] , Z H 0 = H 0 Z , (44) \Omega=
\begin{bmatrix}
Z&0&0&0\\
0&B_0&\Psi&T_0\\
0&\Psi&0&0\\
0&-T_0^T&0&V_0
\end{bmatrix},
\qquad ZH_0=H_0Z, \tag*{(44)} Ω = Z 0 0 0 0 B 0 Ψ − T 0 T 0 Ψ 0 0 0 T 0 0 V 0 , Z H 0 = H 0 Z , (44) with H 0 , H 1 H_0,H_1 H 0 , H 1 nondegenerate, Z , B 0 , Ψ , V 0 Z,B_0,\Psi,V_0 Z , B 0 , Ψ , V 0 skew, and empty blocks allowed.
Proof. The constant linear map L ( Ω 1 ) = Ω 1 H C − C H Ω 1 L(\Omega_1)=\Omega_1HC-CH\Omega_1 L ( Ω 1 ) = Ω 1 H C − C H Ω 1 on skew matrices is injective on a constant complement of its kernel, so the component of Ω \Omega Ω in that complement is polynomial; subtracting it gives (42) . Normalize C = diag ( I r , 0 r ′ ) C=\operatorname{diag}(I_r,0_{r'}) C = diag ( I r , 0 r ′ ) and let n 0 n_0 n 0 be the nullity of H H H on the first r r r -space; split that space orthogonally into its nondegenerate part and radical. The radical pairs with the last r ′ r' r ′ -space with row rank n 0 n_0 n 0 , since H H H is nondegenerate, so n 0 ≤ r ′ n_0\leq r' n 0 ≤ r ′ . Changes [ O T 0 N ] \begin{bmatrix}O&T\\0&N\end{bmatrix} [ O 0 T N ] with O O O orthogonal and N N N invertible preserve C C C : the first d d d rows of T T T remove the cross terms of H 0 H_0 H 0 , N N N normalizes the radical pairing to [ I n 0 , 0 ] [I_{n_0},0] [ I n 0 , 0 ] , and the remaining rows of T T T remove the terms on the paired n 0 n_0 n 0 -space. This gives (43) , and nondegeneracy of H H H gives that of H 1 H_1 H 1 ; comparing blocks in Ω H C = C H Ω \Omega HC=CH\Omega Ω H C = C H Ω gives (44) ; the two copies of Ψ \Psi Ψ have the same sign because Ψ T = − Ψ \Psi^T=-\Psi Ψ T = − Ψ .
Completion of Proposition 15.1. Apply Lemma 15.4 with r = rank Q X r=\operatorname{rank} Q_X r = rank Q X ; the subtracted matrix is polynomial, so Ω \Omega Ω keeps its double pole. Diagonalize H 0 H_0 H 0 orthogonally. Since Z H 0 = H 0 Z ZH_0=H_0Z Z H 0 = H 0 Z by (44) , Z Z Z splits into skew blocks on the nonzero eigenspaces of H 0 H_0 H 0 , each definite for the indefinite form H H H and so of dimension at most seven; the skew block on the last three spaces of (44) has dimension r ′ + n 0 = 8 − d r'+n_0=8-d r ′ + n 0 = 8 − d , also at most seven if d > 0 d>0 d > 0 . These are exact constant block decompositions, so the corresponding columns of P Ω P\Omega P Ω are polynomial products of linear, generically injective column subpencils with the blocks, and Lemma 15.2 excludes a double pole in every block of dimension at most seven. If d > 0 d>0 d > 0 , every block has dimension at most seven, so none carries the double pole of Ω \Omega Ω ; hence d = 0 d=0 d = 0 , that is
C H C = 0 , (45) CHC=0, \tag*{(45)} C H C = 0 , (45)
so the range of C C C is totally H H H -isotropic and r r r is at most the Witt index of H H H ; hence r ≤ 4 r\leq4 r ≤ 4 .
Ranks two and three. Since Q X Q_X Q X is real irreducible and semidefinite, r ≥ 2 r\geq2 r ≥ 2 . With d = 0 d=0 d = 0 we have n 0 = r n_0=r n 0 = r , ℓ = 8 − 2 r \ell=8-2r ℓ = 8 − 2 r and
Ω = [ B 0 Ψ T 0 Ψ 0 0 − T 0 T 0 V 0 ] . \Omega=\begin{bmatrix}
B_0 & \Psi& T_0\\
\Psi& 0 & 0\\
-T_0^T & 0 & V_0
\end{bmatrix}. Ω = B 0 Ψ − T 0 T Ψ 0 0 T 0 0 V 0 .
Write P = [ P 2 , P 3 , P 4 ] P=[P_2,P_3,P_4] P = [ P 2 , P 3 , P 4 ] by the column blocks on U 2 , U 3 , U 4 U_2,U_3,U_4 U 2 , U 3 , U 4 ; the middle columns of P Ω P\Omega P Ω are P 2 Ψ P_2\Psi P 2 Ψ . For r = 2 , 3 r=2,3 r = 2 , 3 a pole of the skew block Ψ \Psi Ψ , even a simple one, would need r ≥ 4 r\geq4 r ≥ 4 by Lemma 15.2, and all its denominators divide Q X 2 Q_X^2 Q X 2 , so Ψ \Psi Ψ is polynomial. Subtracting it leaves the double pole on the first and last spaces, of total dimension 8 − r < 8 8-r<8 8 − r < 8 , which the same lemma excludes.
Rank four. Now H H H has signature ( 4 , 4 ) (4,4) ( 4 , 4 ) and the normalization gives
H ′ = Λ ′ = [ 0 I 4 I 4 0 ] , C ′ = [ I 4 0 0 0 ] , Ω ′ = [ B 0 Ψ Ψ 0 ] , P ′ = [ P 2 , P 3 ] . H'=\Lambda'=\begin{bmatrix}
0&I_4\\
I_4&0
\end{bmatrix},\qquad
C'=\begin{bmatrix}
I_4&0\\
0&0
\end{bmatrix},\qquad
\Omega'=\begin{bmatrix}
B_0&\Psi\\
\Psi&0
\end{bmatrix},\qquad
P'=[P_2,P_3]. H ′ = Λ ′ = [ 0 I 4 I 4 0 ] , C ′ = [ I 4 0 0 0 ] , Ω ′ = [ B 0 Ψ Ψ 0 ] , P ′ = [ P 2 , P 3 ] .
Polynomiality of the four columns P 2 Ψ P_2\Psi P 2 Ψ excludes a double pole of Ψ \Psi Ψ ; if Ψ \Psi Ψ were polynomial, subtracting it would leave a double pole in a four-dimensional block, again excluded. Thus Ψ \Psi Ψ has a simple Q X Q_X Q X -pole, so every maximal four-minor of P 2 P_2 P 2 is divisible by Q X 2 Q_X^2 Q X 2 . These minors have degree four, so they are constant multiples of Q X 2 Q_X^2 Q X 2 , not all zero; their Plücker vector gives one fixed real four-space T T T with Im P 2 ( Y ) ⊂ T \operatorname{Im}P_2(Y)\subset T Im P 2 ( Y ) ⊂ T , first generically and then for all Y Y Y by polynomial continuation. Choose 0 ≠ η ∈ T ⊥ 0\ne\eta\in T^\perp 0 = η ∈ T ⊥ in the original Euclidean lower output coordinates and put z = K X η z=K_X\eta z = K X η . Only coefficient coordinates were changed, so η T P ′ ( Y ) = ( 0 , b ( Y ) T ) \eta^TP'(Y)=(0,b(Y)^T) η T P ′ ( Y ) = ( 0 , b ( Y ) T ) and P ′ Λ ′ P ′ T = P Λ P T P'\Lambda'P'^T=P\Lambda P^T P ′ Λ ′ P ′ T = P Λ P T . The hyperbolic Λ ′ \Lambda' Λ ′ gives ⟨ J t X + Y z , z ⟩ = η T P ′ ( Y ) Λ ′ P ′ ( Y ) T η = 0 \langle J_{tX+Y}z,z\rangle=\eta^TP'(Y)\Lambda'P'(Y)^T\eta=0 ⟨ J tX + Y z , z ⟩ = η T P ′ ( Y ) Λ ′ P ′ ( Y ) T η = 0 for all real t , Y t,Y t , Y ; these are polynomial identities, so no direction is lost at the frame singularities. The symmetry ⟨ J z V , V ⟩ = ⟨ J V z , z ⟩ \langle J_zV,V\rangle=\langle J_Vz,z\rangle ⟨ J z V , V ⟩ = ⟨ J V z , z ⟩ of the curvature tensor gives ⟨ J z V , V ⟩ = 0 \langle J_zV,V\rangle=0 ⟨ J z V , V ⟩ = 0 for every real V V V , and polarization gives J z = 0 J_z=0 J z = 0 , contradicting rank eight at z ≠ 0 z\ne0 z = 0 . So each of the ranks r = 2 , 3 , 4 r=2,3,4 r = 2 , 3 , 4 is excluded, which proves Proposition 15.1. □
Closed kernels and a fixed ten-dimensional image
Throughout this section J x \mathcal{J}_x J x is the Jacobi family of a shifted Osserman algebraic curvature tensor of rank eight on V = R j V=\mathbb{R}^j V = R j , q ( x ) = ⟨ x , x ⟩ q(x)=\langle x,x\rangle q ( x ) = ⟨ x , x ⟩ , polynomial identities are extended complex linearly, and the distinct nonzero reduced eigenvalues are λ 1 q ( x ) , … , λ m q ( x ) \lambda_1q(x),\ldots,\lambda_mq(x) λ 1 q ( x ) , … , λ m q ( x ) . For a real anchor x ≠ 0 x\ne0 x = 0 put E x = Im J x E_x=\operatorname{Im}\mathcal{J}_x E x = Im J x and K x = ker J x K_x=\ker\mathcal{J}_x K x = ker J x . At a regular real unit anchor e e e (written X X X in the other sections), in orthonormal image and kernel frames M e , K e M_e,K_e M e , K e , the coherent factor at z = t e + Y z=te+Y z = t e + Y , Y ⊥ e Y\perp e Y ⊥ e , is
M z = [ t I + F ( Y ) P ( Y ) ] , J z = M z Λ M z T , (46) M_z=\begin{bmatrix}
tI+F(Y)\\
P(Y)
\end{bmatrix},\qquad
\mathcal{J}_z=M_z\Lambda M_z^T, \tag*{(46)} M z = [ t I + F ( Y ) P ( Y ) ] , J z = M z Λ M z T , (46)
with Λ \Lambda Λ invertible constant diagonal, F F F rational, and P P P a generically invertible linear 8 × 8 8\times8 8 × 8 matrix.
Lemma 16.1 (Closed-kernel criterion). Suppose that, on a nonempty Euclidean-open set of real anchors x x x ,
J y z = 0 for every y , z ∈ K x . (47) \mathcal{J}_{yz}=0 \qquad\text{for every }y,z\in\mathcal{K}_x. \tag*{(47)} J y z = 0 for every y , z ∈ K x . (47) Then J x 2 = 0 \mathcal{J}_x^2=0 J x 2 = 0 for every complex x x x with q ( x ) = 0 q(x)=0 q ( x ) = 0 , and at every regular real anchor e e e the full coherent-partner Gram matrix is radial: M t e + Y T M t e + Y = ( t 2 + q ( Y ) ) I M_{te+Y}^{T}M_{te+Y}=(t^2+q(Y))I M t e + Y T M t e + Y = ( t 2 + q ( Y )) I as a rational identity.
Proof. Over C ( Y ) \mathbb{C}(Y) C ( Y ) adjoin α \alpha α with α 2 = − q ( Y ) \alpha^2=-q(Y) α 2 = − q ( Y ) ; as q ( Y ) q(Y) q ( Y ) has rank fifteen it is not a square up to sign there. The lower block P ( Y ) P(Y) P ( Y ) stays invertible, so M α e + Y M_{\alpha e+Y} M α e + Y is injective and its transpose surjective, and, Λ \Lambda Λ being invertible, J α e + Y = M α e + Y Λ M α e + Y T \mathcal{J}_{\alpha e+Y}=M_{\alpha e+Y}\Lambda M_{\alpha e+Y}^{T} J α e + Y = M α e + Y Λ M α e + Y T has rank eight. Hence the complex null quadric contains a dense set of rank-eight Jacobi operators.
Every nine-minor of J x \mathcal{J}_x J x vanishes identically. On the rank-eight locus choose a rational kernel basis using a nonzero eight-minor; substituting linear combinations of it for y , z y,z y , z in (47) gives finitely many rational identities in x x x , which vanish on a nonempty real open set and hence identically. These minor charts cover the generic null locus, so (47) holds at its generic complex points.
Fix such a null point x x x and set K = ker J x \mathcal{K}=\ker\mathcal{J}_x K = ker J x , so dim K = 8 \dim\mathcal{K}=8 dim K = 8 and J x \mathcal{J}_x J x is nilpotent by the Osserman characteristic polynomial. If q ∣ K ≠ 0 q|_{\mathcal{K}}\ne0 q ∣ K = 0 , choose y ∈ K y\in\mathcal{K} y ∈ K with q ( y ) ≠ 0 q(y)\ne0 q ( y ) = 0 . Then (47) , continued above, gives J y K = 0 \mathcal{J}_y\mathcal{K}=0 J y K = 0 , and J y ∏ j = 1 m ( J y − λ j q ( y ) I ) = 0 \mathcal{J}_y\prod_{j=1}^{m}(\mathcal{J}_y-\lambda_jq(y)I)=0 J y ∏ j = 1 m ( J y − λ j q ( y ) I ) = 0 holds over C \mathbb{C} C by continuation; its roots are distinct because q ( y ) ≠ 0 q(y)\ne0 q ( y ) = 0 , so J y \mathcal{J}_y J y is diagonalizable; its characteristic polynomial gives rank eight, whence K = ker J y \mathcal{K}=\ker\mathcal{J}_y K = ker J y . The kernel of a self-adjoint diagonalizable operator on a nondegenerate bilinear space is nondegenerate, so K \mathcal{K} K is nondegenerate, and self-adjointness of J x \mathcal{J}_x J x gives Im J x = K ⊥ \operatorname{Im}\mathcal{J}_x=\mathcal{K}^{\perp} Im J x = K ⊥ with J x \mathcal{J}_x J x invertible on K ⊥ \mathcal{K}^{\perp} K ⊥ , contradicting nilpotency. Therefore q ∣ K = 0 q|_{\mathcal{K}}=0 q ∣ K = 0 ; by polarization K \mathcal{K} K is totally isotropic, and dim K = 8 \dim\mathcal{K}=8 dim K = 8 gives K = K ⊥ \mathcal{K}=\mathcal{K}^{\perp} K = K ⊥ , so Im J x = K ⊥ = ker J x \operatorname{Im}\mathcal{J}_x=\mathcal{K}^{\perp}=\ker\mathcal{J}_x Im J x = K ⊥ = ker J x and J x 2 = 0 \mathcal{J}_x^2=0 J x 2 = 0 . The null quadric is irreducible, so this extends to all its points, including those of smaller Jacobi rank.
Put Φ ( z ) = M z T M z \Phi(z)=M_z^TM_z Φ ( z ) = M z T M z . At z = α e + Y z=\alpha e+Y z = α e + Y , 0 = J z 2 = M z Λ Φ ( z ) Λ M z T 0=\mathcal{J}_z^2=M_z\Lambda\Phi(z)\Lambda M_z^T 0 = J z 2 = M z ΛΦ ( z ) Λ M z T gives Φ ( z ) = 0 \Phi(z)=0 Φ ( z ) = 0 by left and right inverses. With Φ ( t e + Y ) = t 2 I + t S ( Y ) + Φ ( Y ) \Phi(te+Y)=t^2I+tS(Y)+\Phi(Y) Φ ( t e + Y ) = t 2 I + tS ( Y ) + Φ ( Y ) this reads α S ( Y ) + Φ ( Y ) − q ( Y ) I = 0 \alpha S(Y)+\Phi(Y)-q(Y)I=0 α S ( Y ) + Φ ( Y ) − q ( Y ) I = 0 , and independence of 1 , α 1,\alpha 1 , α over C ( Y ) \mathbb{C}(Y) C ( Y ) gives S = 0 S=0 S = 0 and Φ ( Y ) = q ( Y ) I \Phi(Y)=q(Y)I Φ ( Y ) = q ( Y ) I .
Lemma 16.2. Let e e e be a real anchor with transpose pencil B ( a ) g = N g a B(a)g=N_ga B ( a ) g = N g a as in (28) . If Im B ( a ) ⊂ E e \operatorname{Im}B(a)\subset E_e Im B ( a ) ⊂ E e for every real a a a , then (47) holds at e e e .
Proof. By (29) , J K e g = N g Λ N g T \mathcal{J}_{K_eg}=N_g\Lambda N_g^T J K e g = N g Λ N g T , and Im N g ⊂ E e \operatorname{Im}N_g\subset E_e Im N g ⊂ E e by hypothesis, so J K e g \mathcal{J}_{K_eg} J K e g annihilates E e ⊥ = K e E_e^\perp=K_e E e ⊥ = K e for every g g g .
Lemma 16.3 (Fixed ten-dimensional image). At a regular real anchor e e e , let B ( a ) B(a) B ( a ) be the transpose pencil and G ( a ) = B ( a ) B ( a ) T G(a)=B(a)B(a)^T G ( a ) = B ( a ) B ( a ) T , positive semidefinite of rank eight for real a ≠ 0 a\ne0 a = 0 . Suppose that all images of B ( a ) B(a) B ( a ) lie in a fixed real subspace T T T with dim T ≤ 10 \dim T\le10 dim T ≤ 10 , that a possibly reducible projective variety A ⊂ P C 7 A\subset\mathbb{P}_{\mathbb{C}}^7 A ⊂ P C 7 is a reduced complete intersection A = V ( h , q 0 ) A=V(h,q_0) A = V ( h , q 0 ) of independent real quadrics with q 0 q_0 q 0 positive definite, that G G G has generic rank two on each component of A A A , and that no nonzero real linear form vanishes on any component. Then Im B ( a ) = E e \operatorname{Im}B(a)=E_e Im B ( a ) = E e for every real a ≠ 0 a\ne0 a = 0 .
Proof. The radial column B ( a ) g e = − M e a B(a)g_e=-M_ea B ( a ) g e = − M e a gives E e ⊂ T E_e\subset T E e ⊂ T , so the lower projection Γ ( a ) = K e T B ( a ) \Gamma(a)=K_e^TB(a) Γ ( a ) = K e T B ( a ) is skew with image in a fixed real space W W W , dim W ≤ 2 \dim W\le2 dim W ≤ 2 . If dim W ≤ 1 \dim W\le1 dim W ≤ 1 , then skewness gives Γ = 0 \Gamma=0 Γ = 0 , and rank eight gives the claim. Otherwise Γ ( a ) = ℓ ( a ) J \Gamma(a)=\ell(a)J Γ ( a ) = ℓ ( a ) J for a fixed rank-two skew J J J and a real linear form ℓ \ell ℓ ; if ℓ = 0 \ell=0 ℓ = 0 , rank eight gives the claim as before, so let ℓ ≠ 0 \ell\ne0 ℓ = 0 . On E e ⊕ W E_e\oplus W E e ⊕ W ,
G ( a ) = [ A 0 ( a ) ℓ ( a ) Θ ( a ) ℓ ( a ) Θ ( a ) T ℓ ( a ) 2 D ] G(a)=
\begin{bmatrix}
A_0(a) & \ell(a)\Theta(a)\\
\ell(a)\Theta(a)^T & \ell(a)^2D
\end{bmatrix} G ( a ) = [ A 0 ( a ) ℓ ( a ) Θ ( a ) T ℓ ( a ) Θ ( a ) ℓ ( a ) 2 D ] with D D D constant positive definite of size two and Θ \Theta Θ linear (the degrees follow from the linearity of B B B and Γ = ℓ J \Gamma=\ell J Γ = ℓ J ). By the component hypothesis ℓ ≠ 0 \ell\ne0 ℓ = 0 generically on each component of A A A , where
the lower block has rank two; rank two of G G G makes the Schur complement Ψ = A 0 − Θ D − 1 Θ T \Psi= A_0 - \Theta D^{-1}\Theta^{\mathsf{T}} Ψ = A 0 − Θ D − 1 Θ T vanish there. The degree-two ideal of the reduced complete intersection is spanned by h h h , q 0 q_0 q 0 , so Ψ = h Ψ 0 + q 0 Ψ 1 \Psi= h\Psi_0 + q_0\Psi_1 Ψ = h Ψ 0 + q 0 Ψ 1 with constant real symmetric Ψ 0 , Ψ 1 \Psi_0,\Psi_1 Ψ 0 , Ψ 1 .
At real points with ℓ ≠ 0 \ell\ne0 ℓ = 0 the Schur complement is positive semidefinite of rank six; by continuity Ψ ⪰ 0 \Psi\succeq0 Ψ ⪰ 0 of rank at most six everywhere. At real nonzero points of ker ℓ \ker\ell ker ℓ the image of B B B lies in E e E_e E e , so A 0 A_0 A 0 is positive definite and, as Θ D − 1 Θ T \Theta D^{-1}\Theta^{\mathsf{T}} Θ D − 1 Θ T has rank at most two, Ψ \Psi Ψ has rank at least six. Thus Ψ \Psi Ψ has rank six at every real a ≠ 0 a \ne0 a = 0 . Normalize q 0 = 1 q_0 = 1 q 0 = 1 : the range of h h h is a nontrivial closed interval, on which t Ψ 0 + Ψ 1 ⪰ 0 t\Psi_0 + \Psi_1 \succeq0 t Ψ 0 + Ψ 1 ⪰ 0 has rank six, so by Lemma 12.1 all Ψ ( a ) \Psi(a) Ψ ( a ) have one image F ⊂ E e F \subset E_e F ⊂ E e of dimension six.
On the seven-dimensional real space ker ℓ \ker\ell ker ℓ , compressing A 0 = Ψ + Θ D − 1 Θ T A_0 = \Psi+ \Theta D^{-1}\Theta^{\mathsf{T}} A 0 = Ψ + Θ D − 1 Θ T to the two-dimensional space F ⊥ ∩ E e F^\perp\cap E_e F ⊥ ∩ E e gives Θ 2 ( a ) D − 1 Θ 2 ( a ) T \Theta_2(a)D^{-1}\Theta_2(a)^{\mathsf{T}} Θ 2 ( a ) D − 1 Θ 2 ( a ) T , positive definite for every nonzero a ∈ ker ℓ a \in\ker\ell a ∈ ker ℓ . So the real 2 × 2 2 \times2 2 × 2 matrix Θ 2 ( a ) \Theta_2(a) Θ 2 ( a ) is invertible there; but Θ 2 \Theta_2 Θ 2 is linear from a seven-dimensional space to a four-dimensional matrix space and has a nonzero kernel. So dim W = 2 \dim W = 2 dim W = 2 is impossible; hence Γ = 0 \Gamma= 0 Γ = 0 , and rank eight proves the claim. □
The lemma is applied in Proposition 25.2, with A A A the union of the two conjugate residue images of § 25 .
Small supports and the residue-image incidence
We use the normalization of § 10 : ν ∈ { 7 , 8 } \nu\in\{7,8\} ν ∈ { 7 , 8 } , k = 16 − ν k = 16-\nu k = 16 − ν , and the lower pencil P ( Y ) : R ν → R k P(Y) : \mathbb{R}^{\nu} \to\mathbb{R}^{k} P ( Y ) : R ν → R k is linear and generically injective for Y ∈ X ⊥ Y \in X^\perp Y ∈ X ⊥ , extended by P ( X ) = 0 P(X) = 0 P ( X ) = 0 ; N g N_g N g and B ( a ) g = N g a B(a)g = N_g a B ( a ) g = N g a are as in (10.6). By Lemma 10.2 , for fixed real coefficient and output subspaces U , T U,T U , T ,
P ( Y ) U ⊂ T for all Y , U ≠ 0 ⟹ T = R k . (48) P(Y)U \subset T \text{ for all } Y,\quad U \ne0 \quad\Longrightarrow\quad T = \mathbb{R}^{k}. \tag*{(48)} P ( Y ) U ⊂ T for all Y , U = 0 ⟹ T = R k . (48)
Four- and five-dimensional quadratic supports
For a simple pole along a real irreducible polynomial f f f (a quadric or a quartic), let W W W be its nonzero skew residue over the fraction field of R [ Y ] / ( f ) \mathbb{R}[Y]/(f) R [ Y ] / ( f ) . Its real support is the smallest constant real coefficient subspace whose extension to C \mathbb{C} C or to the residue field contains Im W \operatorname{Im} W Im W ; when f f f splits over C \mathbb{C} C , it contains the residue images on both geometric components.
Lemma 17.1. A simple quadratic pole component with rank-two residue cannot have real support of dimension at most five, whatever other polar factors are present.
Proof. Let the support lie in a fixed real d d d -space U U U . At the prime ( Q ) (Q) ( Q ) the kernel of P ∣ U P|_U P ∣ U contains the rank-two residue image, so the rectangular Smith form has at least two positive valuations and every maximal d d d -minor of P ∣ U P|_U P ∣ U is divisible by Q 2 Q^2 Q 2 . Some such minor is nonzero by generic injectivity, and has degree d d d ; thus d ≥ 4 d \ge4 d ≥ 4 .
If d = 4 d = 4 d = 4 , all maximal minors are constant multiples of Q 2 Q^2 Q 2 , so their Plücker vector has a constant direction and P ( Y ) U P(Y)U P ( Y ) U is one fixed real four-space, first generically, and then, the projection to a complement being polynomial, for every Y Y Y . This contradicts (48) , as k ≥ 8 k \ge8 k ≥ 8 .
If d = 5 d = 5 d = 5 , the maximal minors give
⋀ 5 ( P ( Y ) ∣ U ) = Q ( Y ) 2 ℓ ( Y ) (49) \bigwedge\nolimits^{5}(P(Y)|_U) = Q(Y)^2\ell(Y) \tag*{(49)} ⋀ 5 ( P ( Y ) ∣ U ) = Q ( Y ) 2 ℓ ( Y ) (49)
with ℓ \ell ℓ a nonzero real linear map into ⋀ 5 R k \bigwedge^{5}\mathbb{R}^{k} ⋀ 5 R k . Its generic values are decomposable, and, the Plücker relations being polynomial, so are all its values. Two independent decomposable five-vectors with decomposable sum span five-spaces meeting in dimension four. Fix two members, with intersection T 0 T_0 T 0 and sum T 1 T_1 T 1 : a third member meeting both in dimension four contains T 0 T_0 T 0 or lies in T 1 T_1 T 1 , and a member containing T 0 T_0 T 0 but not in T 1 T_1 T 1 meets a member in T 1 T_1 T 1 not containing T 0 T_0 T 0 in dimension at most three, whereas any two members meet in dimension four. So the five-spaces are constant, or all contain the four-space T 0 T_0 T 0 , or all lie in the six-space T 1 T_1 T 1 . The first and last contradict (17.1). In the second, P ∣ U P|_U P ∣ U modulo T 0 T_0 T 0 is a linear family of matrices of rank at most one, generically and then everywhere by its two-minors. By Lemma 12.6 either all P ( Y ) U P(Y)U P ( Y ) U lie in a fixed five-space, or a fixed four-space U 0 ⊂ U U_0\subset U U 0 ⊂ U has P ( Y ) U 0 ⊂ T 0 P(Y)U_0\subset T_0 P ( Y ) U 0 ⊂ T 0 for all Y Y Y ; both violate (17.1).
Proposition 17.2 (Isolated skew corrections). Let U ⊂ R ν U\subset\mathbb{R}^{\nu} U ⊂ R ν be a fixed real coefficient space of dimension at most five, and write P U = P ∣ U P_U=P|_U P U = P ∣ U in a fixed orthonormal basis of U U U . If Γ ( Y ) \Gamma(Y) Γ ( Y ) is a rational skew-symmetric endomorphism of U U U and P U ( Y ) Γ ( Y ) P_U(Y)\Gamma(Y) P U ( Y ) Γ ( Y ) is polynomial, then Γ \Gamma Γ has no pole along an irreducible real quadratic.
Proof. If Γ \Gamma Γ had a pole of order e > 0 e>0 e > 0 at a quadratic Q Q Q , its leading coefficient W = ( Q e Γ ) m o d Q W=(Q^e\Gamma)\bmod Q W = ( Q e Γ ) mod Q would be nonzero and skew, hence of rank at least two, and polynomiality of P U Γ P_U\Gamma P U Γ gives P U W = 0 P_UW=0 P U W = 0 modulo Q Q Q . So P U P_U P U has corank at least two there and every maximal minor is divisible by Q 2 Q^2 Q 2 . The degree count, the constant-output case and the Plücker argument for d = 5 d=5 d = 5 in the proof of Lemma 17.1 then apply unchanged, each giving a nonzero fixed coefficient subspace mapped by P P P into a proper fixed output space, contrary to (17.1).
A real-irreducible quartic has full real support
Lemma 17.3. For ν = 8 \nu=8 ν = 8 , a simple real-irreducible quartic pole has real residue support of dimension eight, also when it splits into two conjugate complex quadratics.
Proof. Let the quartic be f f f . Lemma 10.4 and deg det P = 8 \deg\det P=8 deg det P = 8 give det P = c f 2 \det P=cf^2 det P = c f 2 with c ≠ 0 c\ne0 c = 0 ; the nonzero simple skew residue has positive even rank, and the same budget forces rank two and Smith exponents ( 1 , 1 , 0 6 ) (1,1,0^6) ( 1 , 1 , 0 6 ) at ( f ) (f) ( f ) . Every seven-minor is divisible by f f f , so P ♯ = adj P / ( c f ) P^\sharp=\operatorname{adj}P/(cf) P ♯ = adj P / ( c f ) is a homogeneous cubic polynomial matrix with P ♯ P = P P ♯ = f I P^\sharp P=PP^\sharp=fI P ♯ P = P P ♯ = f I , and modulo f f f its image is the two-dimensional kernel of P P P , which contains the image of the rank-two residue W W W and therefore equals it. If a constant real row b T b^T b T annihilates the real support of W W W , it annihilates P ♯ P^\sharp P ♯ modulo f f f ; the entries of b T P ♯ b^TP^\sharp b T P ♯ have degree three < deg f <\deg f < deg f , so b T P ♯ = 0 b^TP^\sharp=0 b T P ♯ = 0 , and multiplying by P P P gives f b T = 0 fb^T=0 f b T = 0 , so b = 0 b=0 b = 0 .
Dimension and rank of the closed image
From now on ν = k = 8 \nu=k=8 ν = k = 8 and H = Λ − 1 H=\Lambda^{-1} H = Λ − 1 is indefinite. Put h ( a ) = a T H a h(a)=a^THa h ( a ) = a T H a and G ( a ) = B ( a ) B ( a ) T G(a)=B(a)B(a)^T G ( a ) = B ( a ) B ( a ) T . Fix one irreducible complex pole hypersurface Z ⊂ P ( X C ⊥ ) = P 14 Z\subset\mathbb{P}(X^\perp_{\mathbb{C}})=\mathbb{P}^{14} Z ⊂ P ( X C ⊥ ) = P 14 with generic residue rank two, let I \mathcal{I} I be the closure of its projectivized residue-image bundle in Z × P 7 Z\times\mathbb{P}^7 Z × P 7 , and let A A A be the image of I \mathcal{I} I in P 7 \mathbb{P}^7 P 7 . Both are irreducible, dim I = 13 + 1 = 14 \dim\mathcal{I}=13+1=14 dim I = 13 + 1 = 14 (the fibres over Z Z Z are projective lines); write α = dim A \alpha=\dim A α = dim A and let s s s be the generic rank of B ( a ) B(a) B ( a ) on A A A ; we say that the pole component has type ( α , s ) (\alpha,s) ( α , s ) .
Lemma 17.4. One has 4 ≤ s ≤ α ≤ 5 4\le s\le\alpha\le5 4 ≤ s ≤ α ≤ 5 . Moreover A A A has no real point, A ⊂ V ( h ) A\subset V(h) A ⊂ V ( h ) , and rank G ( a ) = 2 rank B ( a ) − 8 \operatorname{rank}G(a)=2\operatorname{rank}B(a)-8 rank G ( a ) = 2 rank B ( a ) − 8 for every nonzero complex a a a .
Proof. The rank identity and s ≥ 4 s\ge4 s ≥ 4 are Lemma 10.3. The residue identities (10.9) give A ⊂ V ( h ) A\subset V(h) A ⊂ V ( h ) , so α ≤ 6 \alpha\le6 α ≤ 6 , and put every fibre of I → A \mathcal{I}\to A I → A inside L a = { [ Y ] ∈ P 14 : P ( Y ) a = 0 } L_a=\{[Y]\in\mathbb{P}^{14}:P(Y)a=0\} L a = {[ Y ] ∈ P 14 : P ( Y ) a = 0 } . The map Y ↦ P ( Y ) a Y\mapsto P(Y)a Y ↦ P ( Y ) a has transpose B ( a ) B(a) B ( a ) (as P ( X ) = 0 P(X)=0 P ( X ) = 0 ), so generically dim L a = 14 − s \dim L_a=14-s dim L a = 14 − s ; the generic fibre has dimension 14 − α 14-\alpha 14 − α , whence s ≤ α s\le\alpha s ≤ α . The minors giving rank B ≤ α B\le\alpha B ≤ α vanish generically on A A A , hence on all of A A A ; a nonzero real coefficient gives rank eight (Lemma 10.2), so A A A has no real point. If α = 6 \alpha=6 α = 6 , A A A would be the irreducible sixfold V ( h ) V(h) V ( h ) , which has real points since h h h is indefinite. Hence α ≤ 5 \alpha\le5 α ≤ 5 .
Full real support with α = 5 \alpha= 5 α = 5 and s = 4 s = 4 s = 4
Proposition 17.5. Let a simple irreducible real quadratic or quartic pole have total real residue support eight. If a complex residue-image component has α = 5 \alpha= 5 α = 5 and s = 4 s = 4 s = 4 , then Im B ( a ) = E X \operatorname{Im} B(a) = E_X Im B ( a ) = E X for every real a ≠ 0 a \ne0 a = 0 .
Proof. The rank identity makes G = 0 G = 0 G = 0 on the image A A A . If the real pole polynomial is geometrically irreducible, conjugation preserves its unique component, so A A A is real-defined, and its span, a real-defined space containing the residue image, is P 7 \mathbb{P}^{7} P 7 by total support eight. For such A A A the bound deg A ≥ 1 + codim A \deg A \ge1 + \operatorname{codim} A deg A ≥ 1 + codim A for nondegenerate varieties [10 p. 1, (*)] gives deg A ≥ 3 \deg A \ge3 deg A ≥ 3 , and a generic transverse real complementary plane meets the real-pointless A A A in conjugate pairs, so deg A ≥ 4 \deg A \ge4 deg A ≥ 4 . If every entry of G G G is a multiple of h h h , evaluation at a nonzero real h h h -null vector contradicts rank eight. Otherwise choose a real entry f f f independent of h h h . The complete intersection V ( h , f ) V(h,f) V ( h , f ) has degree four and contains A A A ; being Cohen–Macaulay it is unmixed, and since deg V ( h , f ) = 4 ≤ deg A \deg V(h,f) = 4 \le\deg A deg V ( h , f ) = 4 ≤ deg A , A A A is its only component, with multiplicity one, so ( h , f ) (h,f) ( h , f ) is generically reduced and unmixed, hence radical, and its degree-two part is span ( h , f ) \operatorname{span}(h,f) span ( h , f ) . Lemma 12.2 and Remark 12.3 give the claim.
Otherwise the real pole polynomial, being real irreducible, has exactly two conjugate irreducible factors, with images A A A , A ‾ \overline{A} A spanning U U U , U ‾ \overline{U} U ; total real support eight means U + U ‾ = C 8 U + \overline{U} = \mathbb{C}^{8} U + U = C 8 . Both images are fivefolds on which G G G vanishes. Put d = dim C U d = \dim_{\mathbb{C}} U d = dim C U . If d ≤ 6 d \le6 d ≤ 6 , then, A A A being a fivefold, d = 6 d = 6 d = 6 and A = P ( U ) A = \mathbb{P}(U) A = P ( U ) , and U ∩ U ‾ U \cap\overline{U} U ∩ U is conjugation-invariant of dimension at least four, so it contains a real point of A A A , contrary to Lemma 17.4. If d = 7 d = 7 d = 7 , h ∣ U h|_{U} h ∣ U has rank at least six and its irreducible quadric fivefold is A A A ; the spans are distinct hyperplanes ℓ = 0 \ell= 0 ℓ = 0 , ℓ ‾ = 0 \overline{\ell} = 0 ℓ = 0 , and the two quadrics exhaust the degree-four complete intersection V ( h , ℓ ℓ ‾ ) V(h,\ell\overline{\ell}) V ( h , ℓ ℓ ) , their reduced union by unmixedness, whose quadratic ideal has scalar span two; Lemma 12.2 and Remark 12.3 apply again. If d = 8 d = 8 d = 8 and A = A ‾ A = \overline{A} A = A , the first paragraph applies; if d = 8 d = 8 d = 8 and they are distinct, each has degree at least three [10 p. 1, (*)] , so an entry f f f independent of h h h would put a union of degree at least six inside V ( h , f ) V(h,f) V ( h , f ) of degree four, and all entries are multiples of h h h , already excluded.
On a nonempty real open set of anchors this gives (47) by Lemma 16.2 , so Lemma 16.1 applies, and Proposition 9.1 (2) makes R R R Clifford (§ 27 ).
Saturated fibres
Lemma 17.6. For the incidence above, suppose α = s \alpha= s α = s . For generic a ∈ A a \in A a ∈ A the fibre I a \mathcal{I}_a I a is the whole projective linear space L a L_a L a . Consequently two distinct complex pole hypersurfaces with this equality cannot have the same residue image.
Proof. The closed fibre I a ⊂ L a \mathcal{I}_a \subset L_a I a ⊂ L a has dimension 14 − α = 14 − s = dim L a 14 - \alpha= 14 - s = \dim L_a 14 − α = 14 − s = dim L a at a generic point, and a proper closed subset of an irreducible projective linear space has smaller dimension. If two pole incidences had the same image and satisfied the equality, their fibres would agree over a common dense open subset of A A A , so their irreducible closures, and their projections to P 14 \mathbb{P}^{14} P 14 , would agree, and the two pole hypersurfaces would coincide.
Aligned quadratic families and two quadratic poles
We work at Jacobi rank eight and, as in the proof of Theorem A in § 27 , assume throughout this section and § 19 that R R R is not Clifford. We consider the eleven patterns with a reduced eigenvalue of multiplicity seven that Theorems C, E and F leave open, namely ( 7 , 5 , 3 ) (7,5,3) ( 7 , 5 , 3 ) , ( 7 , 4 , 4 ) (7,4,4) ( 7 , 4 , 4 ) , ( 7 , 4 , 3 , 1 ) (7,4,3,1) ( 7 , 4 , 3 , 1 ) , ( 7 , 4 , 2 , 2 ) (7,4,2,2) ( 7 , 4 , 2 , 2 ) , ( 7 , 4 , 2 , 1 , 1 ) (7,4,2,1,1) ( 7 , 4 , 2 , 1 , 1 ) , ( 7 , 4 , 1 4 ) (7,4,1^{4}) ( 7 , 4 , 1 4 ) , ( 7 , 3 , 3 , 2 ) (7,3,3,2) ( 7 , 3 , 3 , 2 ) , ( 7 , 3 , 3 , 1 , 1 ) (7,3,3,1,1) ( 7 , 3 , 3 , 1 , 1 ) , ( 7 , 3 , 2 , 2 , 1 ) (7,3,2,2,1) ( 7 , 3 , 2 , 2 , 1 ) , ( 7 , 3 , 2 , 1 3 ) (7,3,2,1^{3}) ( 7 , 3 , 2 , 1 3 ) and ( 7 , 3 , 1 5 ) (7, 3, 1^{5}) ( 7 , 3 , 1 5 ) , where Λ \Lambda Λ is not a multiple of the identity. At a regular unit anchor X X X , let E = Im J X E = \operatorname{Im} \mathcal{J}_X E = Im J X , K = ker J X K = \ker\mathcal{J}_X K = ker J X , with orthonormal frames M X M_X M X , K X K_X K X ; write Y = M X v + u ∈ X ⊥ Y = M_Xv + u \in X^\perp Y = M X v + u ∈ X ⊥ with u ∈ K ∩ X ⊥ u \in K \cap X^\perp u ∈ K ∩ X ⊥ , and M Y = M X F ( Y ) + K X P ( Y ) M_Y = M_XF(Y) + K_XP(Y) M Y = M X F ( Y ) + K X P ( Y ) , H = Λ − 1 H = \Lambda^{-1} H = Λ − 1 . The lower pencil P P P is linear and generically invertible with radial row − v -v − v ; the transpose pencil B ( a ) B(a) B ( a ) , B ( a ) T Y = P ( Y ) a B(a)^{T}Y = P(Y)a B ( a ) T Y = P ( Y ) a , is as in § 12 . We use the cross matrix T ( X , Y ) T(X,Y) T ( X , Y ) and its reduced denominator D ( X , Y ) D(X,Y) D ( X , Y ) of Proposition 11.4 : D D D is bihomogeneous, shear invariant and has no factor in one slot only, and if D D D is quadratic in each slot, it is a nonnegative symmetric biquadratic sectional form D ( X , Y ) = ⟨ A X Y , Y ⟩ D(X,Y) = \langle A_XY,Y\rangle D ( X , Y ) = ⟨ A X Y , Y ⟩ of an auxiliary curvature family A A A . Branch arguments take place on a nonempty real open anchor region as in Lemma 27.1 ; a branch is sole-quadratic if its reduced denominator is one real quadric Q Q Q and two-quadratic if it is Q 1 Q 2 Q_1Q_2 Q 1 Q 2 . Besides an aligned transpose pencil (§ 10 ), we call a semidefinite quadric Q ( X , ⋅ ) Q(X,\mathord\cdot) Q ( X , ⋅ ) aligned if it vanishes on ker J X \ker\mathcal{J}_X ker J X , and a quadratic family A A A aligned if Im A X ⊂ E X \operatorname{Im} A_X \subset E_X Im A X ⊂ E X .
A single quadratic denominator
The partners Y Y Y of a fixed real anchor X X X with Im J X ∩ Im J Y = 0 \operatorname{Im} \mathcal{J}_X \cap\operatorname{Im} \mathcal{J}_Y = 0 Im J X ∩ Im J Y = 0 (disjoint pairs) form a nonempty open set: the proof of [23 Lemma 2(1), p. 7 of arXiv:math/0204258v2] fixes an arbitrary unit X X X before proving openness and density in Y Y Y ; and T T T is regular near such a pair.
Lemma 18.1. Suppose the intrinsic denominator is quadratic in each slot. Fix a regular unit X X X and put Q ( Y ) = D ( X , Y ) Q(Y) = D(X,Y) Q ( Y ) = D ( X , Y ) . Suppose Q Q Q is unchanged by translations in a real subspace U ⊂ X ⊥ U \subset X^\perp U ⊂ X ⊥ and the full coherent factor is affine in those directions. Then A V = c ( V ) A X A_V = c(V)A_X A V = c ( V ) A X for V ∈ R X + U V \in\mathbb{R}X + U V ∈ R X + U , where c c c is a positive definite quadratic form with c ( X ) = 1 c(X) = 1 c ( X ) = 1 .
Proof. First, no nonzero real V V V has D ( V , Y ) ≡ 0 D(V,Y) \equiv0 D ( V , Y ) ≡ 0 . Otherwise nonnegativity in the first slot gives D ( V + t Z , Y ) = t 2 D ( Z , Y ) D(V + tZ,Y) = t^2D(Z,Y) D ( V + tZ , Y ) = t 2 D ( Z , Y ) ; with the polynomial matrix T ~ = D T \widetilde{T} = DT T = D T , at a disjoint pair ( V , Y ) (V,Y) ( V , Y ) the coefficient of t 2 t^2 t 2 in T ~ ( V + t Z , Y ) \widetilde{T}(V + tZ,Y) T ( V + tZ , Y ) is D ( Z , Y ) T ( V , Y ) D(Z,Y)T(V,Y) D ( Z , Y ) T ( V , Y ) . An irreducible factor p ( Y ) p(Y) p ( Y ) of a reduced denominator of an entry of T ( V , Y ) T(V,Y) T ( V , Y ) then divides D ( Z , Y ) D(Z,Y) D ( Z , Y ) for Z Z Z in a Zariski-dense set; reducing D D D modulo p p p over the domain R [ Y ] / ( p ) \mathbb{R}[Y]/(p) R [ Y ] / ( p ) gives p ∣ D p \mid D p ∣ D in R [ Z , Y ] \mathbb{R}[Z,Y] R [ Z , Y ] , a factor in one slot only, excluded by Proposition 11.4 . So T ( V , Y ) T(V,Y) T ( V , Y ) is polynomial and homogeneous linear in Y Y Y . Write T ( V , Y ) = M Y ′ Λ M T V T(V,Y) = M'_Y\Lambda M_T^V T ( V , Y ) = M Y ′ Λ M T V with a fixed real factor J V = M V Λ M V T J_V = M_V\Lambda M_V^T J V = M V Λ M V T , as in the proof of Proposition 11.4 . Then M Y ′ = T ( V , Y ) M V ( M V T M V ) − 1 Λ − 1 M'_Y = T(V,Y)M_V(M_V^TM_V)^{-1}\Lambda^{-1} M Y ′ = T ( V , Y ) M V ( M V T M V ) − 1 Λ − 1 is linear in Y Y Y , and J Y = M Y ′ Λ M Y ′ T J_Y = M'_Y\Lambda M_Y^{\prime T} J Y = M Y ′ Λ M Y ′ T on a dense set of Y Y Y , hence for all Y Y Y ; M Y ′ M'_Y M Y ′ is injective for real Y ≠ 0 Y \ne0 Y = 0 because J Y \mathcal{J}_Y J Y has rank eight. So Proposition 9.1[1 ] makes the tensor Clifford, a contradiction.
For V ∈ R X + U V \in\mathbb{R}X + U V ∈ R X + U , affineness gives M Y + t V = M Y + t C V ( Y ) M_{Y+tV} = M_Y + tC_V(Y) M Y + t V = M Y + t C V ( Y ) with C V C_V C V independent of t t t , invariant under translations in R X + U \mathbb{R}X + U R X + U , and rational with poles only on Q Q Q . The quadratic coefficient gives C V Λ C V T = J V C_V\Lambda C_V^T = J_V C V Λ C V T = J V . For a fixed real factor M V 0 M_V^0 M V 0 of J V J_V J V write C V ( Y ) = M V 0 R V ( Y ) C_V(Y) = M_V^0R_V(Y) C V ( Y ) = M V 0 R V ( Y ) with R V Λ R V T = Λ R_V\Lambda R_V^T = \Lambda R V Λ R V T = Λ ; both R V R_V R V and R V − 1 = Λ R V T Λ − 1 R_V^{-1} = \Lambda R_V^T\Lambda^{-1} R V − 1 = Λ R V T Λ − 1 have poles only on Q Q Q , and R V ( Y + t V ) = R V ( Y ) R_V(Y+tV) = R_V(Y) R V ( Y + t V ) = R V ( Y ) . So M Y = M Y R V ( Y ) − 1 M_Y = M_YR_V(Y)^{-1} M Y = M Y R V ( Y ) − 1 satisfies M Y + t V ′ = M Y ′ + t M V 0 M_{Y+tV}' = M_Y' + tM_V^0 M Y + t V ′ = M Y ′ + t M V 0 and M Y ′ Λ M Y ′ T = J Y M_Y'\Lambda M_Y^{\prime T} = J_Y M Y ′ Λ M Y ′ T = J Y , and uniqueness in Proposition 11.4 gives T ( V , Y ) = M Y ′ Λ M V ′ T T(V,Y) = M_Y'\Lambda M_V^{\prime T} T ( V , Y ) = M Y ′ Λ M V ′ T . The factor M ′ M' M ′ is coherent at V V V with poles only on Q Q Q , and it has a pole: otherwise M Y ′ M_Y' M Y ′ , homogeneous of degree one without poles, would be linear in Y Y Y , a global linear factor of J J J injective at every real Y ≠ 0 Y \ne0 Y = 0 , and Proposition 9.1[1 ] would make the tensor Clifford. Its denominator divides the nonzero quadratic D ( V , Y ) D(V,Y) D ( V , Y ) , so D ( V , Y ) = c ( V ) Q ( Y ) D(V,Y) = c(V)Q(Y) D ( V , Y ) = c ( V ) Q ( Y ) ; nonnegativity gives c ( V ) > 0 c(V) > 0 c ( V ) > 0 for V ≠ 0 V \ne0 V = 0 , and evaluation at a Y 0 Y_0 Y 0 with Q ( Y 0 ) > 0 Q(Y_0) > 0 Q ( Y 0 ) > 0 shows that c c c is quadratic. □ \square □
Proposition 18.2. Suppose the intrinsic denominator is one quadric with a simple pole, D ( X , Y ) = ⟨ A X Y , Y ⟩ D(X,Y) = \langle A_XY,Y\rangle D ( X , Y ) = ⟨ A X Y , Y ⟩ . Then A ⪰ 0 A \succeq0 A ⪰ 0 with Im A X = E X \operatorname{Im} A_X = E_X Im A X = E X cannot hold at every anchor X X X of a nonempty real open region.
Proof. Fix a generic regular unit X X X in the region, so that D ( X , ⋅ ) D(X,\mathord\cdot) D ( X , ⋅ ) is the reduced denominator. Image inclusion holds for all z z z by continuation of the spectral projector identities, so Lemma 15.3 gives A z = M z C M z T A_z = M_z C M_z^T A z = M z C M z T with C > 0 C > 0 C > 0 and Q ( Y ) = v T C v Q(Y) = v^T C v Q ( Y ) = v T C v . The inequality λ min ( C ) ∥ M z ∥ F 2 ≤ tr A z \lambda_{\min}(C)\lVert M_z\rVert_F^2 \le\operatorname{tr} A_z λ m i n ( C ) ∥ M z ∥ F 2 ≤ tr A z bounds the partner on the regular real unit sphere. Write F ( v , u ) = n ( v , u ) / Q ( v ) F(v,u) = n(v,u)/Q(v) F ( v , u ) = n ( v , u ) / Q ( v ) with n n n homogeneous cubic; under v ↦ t v v \mapsto tv v ↦ t v its terms of bidegrees ( 0 , 3 ) (0,3) ( 0 , 3 ) and ( 1 , 2 ) (1,2) ( 1 , 2 ) would give poles of orders two and one, so boundedness removes them, and F F F is affine in all seven u u u directions.
Lemma 18.1 gives A V = c ( V ) A X A_V = c(V)A_X A V = c ( V ) A X on K \mathcal{K} K . On the regular locus, Cauchy–Binet and the scaled spectrum give e 8 ( A z ) = det C det ( M z T M z ) = det C ∥ z ∥ 16 e_8(A_z) = \det C\det(M_z^T M_z) = \det C\lVert z\rVert^{16} e 8 ( A z ) = det C det ( M z T M z ) = det C ∥ z ∥ 16 , where e 8 e_8 e 8 is the sum of principal eight-minors; this polynomial identity holds everywhere, and on K \mathcal{K} K it gives c ( V ) 8 = ∥ V ∥ 16 c(V)^8 = \lVert V\rVert^{16} c ( V ) 8 = ∥ V ∥ 16 , so, c c c being positive, c ( V ) = ∥ V ∥ 2 c(V) = \lVert V\rVert^2 c ( V ) = ∥ V ∥ 2 . The symmetry ⟨ A V Y , Y ⟩ = ⟨ A Y V , V ⟩ \langle A_VY,Y\rangle= \langle A_YV,V\rangle ⟨ A V Y , Y ⟩ = ⟨ A Y V , V ⟩ of the sectional form yields P ( Y ) C P ( Y ) T = ( v T C v ) I P(Y)CP(Y)^T = (v^T C v)I P ( Y ) C P ( Y ) T = ( v T C v ) I . At v = 0 v = 0 v = 0 positivity of C C C gives P = 0 P = 0 P = 0 , so P = P ( v ) P = P(v) P = P ( v ) , and P ( v ) T P ( v ) = ( v T C v ) C − 1 P(v)^TP(v) = (v^T C v)C^{-1} P ( v ) T P ( v ) = ( v T C v ) C − 1 for v ≠ 0 v \ne0 v = 0 . At u = 0 u = 0 u = 0 the Gram determinant identity and positivity give
∥ v ∥ 16 = det ( F ( v ) T F ( v ) + ( v T C v ) C − 1 ) ≥ ( v T C v ) 8 det C . \lVert v\rVert^{16} = \det\left(F(v)^T F(v) + (v^T C v)C^{-1}\right) \ge\frac{(v^T C v)^8}{\det C}. ∥ v ∥ 16 = det ( F ( v ) T F ( v ) + ( v T C v ) C − 1 ) ≥ det C ( v T C v ) 8 .
so the largest eigenvalue of C C C is at most their geometric mean, C = γ I C = \gamma I C = γ I , and equality gives F ( v ) = 0 F(v) = 0 F ( v ) = 0 . Affineness writes F ( v , u ) = L u ( v ) F(v,u) = L_u(v) F ( v , u ) = L u ( v ) with L L L linear in u u u and L u ( t v ) = L u ( v ) L_u(tv) = L_u(v) L u ( t v ) = L u ( v ) . At x X + u + t M X v xX + u + tM_Xv x X + u + t M X v the upper frame block is x I + L u ( v ) xI + L_u(v) x I + L u ( v ) and the lower block is t P ( v ) tP(v) tP ( v ) ; the polynomial limit of A = γ M M T A = \gamma MM^T A = γ M M T as t → 0 t \to0 t → 0 gives ( x I + L u ( v ) ) ( x I + L u ( v ) ) T = ( x 2 + ∥ u ∥ 2 ) I (xI + L_u(v))(xI + L_u(v))^T = (x^2 + \lVert u\rVert^2)I ( x I + L u ( v )) ( x I + L u ( v ) ) T = ( x 2 + ∥ u ∥ 2 ) I , and the transpose product is the same. Adding the lower Gram gives the full radial Gram, and Proposition 9.1 (2) makes R R R Clifford, a contradiction.
Two quadratic poles and semidefinite families
Proposition 18.3. Suppose the denominator at rank eight is Q 1 Q 2 Q_1Q_2 Q 1 Q 2 , with distinct real irreducible semidefinite quadrics. There is no nonzero quadratic curvature family A ⪰ 0 A \succeq0 A ⪰ 0 with Im A z ⊂ Im J z \operatorname{Im} A_z \subset\operatorname{Im} J_z Im A z ⊂ Im J z for every real z ≠ 0 z \ne0 z = 0 .
Proof. Lemma 15.3 gives A z = M z C M z T A_z = M_z C M_z^T A z = M z C M z T with C ⪰ 0 C \succeq0 C ⪰ 0 of rank r ≥ 1 r \ge1 r ≥ 1 and Ω H C − C H Ω \Omega HC - CH\Omega Ω H C − C H Ω polynomial, so Lemma 15.4 applies: with r ′ = 8 − r r' = 8-r r ′ = 8 − r and blocks U 1 , U 2 , U 3 , U 4 U_1,U_2,U_3,U_4 U 1 , U 2 , U 3 , U 4 of dimensions r − n 0 , n 0 , n 0 , r ′ − n 0 r-n_0,n_0,n_0,r'-n_0 r − n 0 , n 0 , n 0 , r ′ − n 0 , n 0 ≤ min ( r , r ′ ) ≤ 4 n_0 \le\min(r,r') \le4 n 0 ≤ min ( r , r ′ ) ≤ 4 , the polar part has the form (15.7). Write P 2 P_2 P 2 for the columns of the normalized P P P on U 2 U_2 U 2 ; the third block column of P Ω P\Omega P Ω is P 2 Ψ P_2\Psi P 2 Ψ . This skew block has no pole along a real irreducible quadratic. Indeed, if S 0 S_0 S 0 is the constant, possibly nonorthogonal, normalizing change of coefficients, then P 2 ( Y ) = P ( Y ) S 0 ∣ U 2 P_2(Y) = P(Y)S_0|_{U_2} P 2 ( Y ) = P ( Y ) S 0 ∣ U 2 , so P 2 P_2 P 2 is P P P restricted to the fixed real coefficient space U ′ = S 0 U 2 U' = S_0U_2 U ′ = S 0 U 2 , of dimension n 0 ≤ 4 n_0 \le4 n 0 ≤ 4 . At such a pole the leading Laurent coefficient of Ψ \Psi Ψ is nonzero and skew, hence of rank at least two, and P 2 P_2 P 2 annihilates it modulo the quadratic; so P ∣ U ′ P|_{U'} P ∣ U ′ has corank at least two there, which the proof of Proposition 17.2 excludes: for n 0 ≤ 3 n_0 \le3 n 0 ≤ 3 by the degree count, and for n 0 = 4 n_0 = 4 n 0 = 4 because P ( Y ) U ′ P(Y)U' P ( Y ) U ′ would be one fixed four-space, contrary to (17.1). Subtract the polynomial block. The remaining polar blocks are Z Z Z and the skew block on U 2 + U 4 U_2+U_4 U 2 + U 4 , of dimension r ′ r' r ′ . Diagonalizing H 0 H_0 H 0 splits Z Z Z into blocks on its nonzero eigenspaces, each definite for the indefinite H H H and of dimension at most seven; the last block also has dimension at most seven.
These blocks have disjoint supports of total dimension ( r − n 0 ) + r ′ ≤ 8 (r-n_0)+r' \le8 ( r − n 0 ) + r ′ ≤ 8 , and each product with its column subencil of P P P is polynomial. Proposition 17.2 therefore permits at most one pole-bearing block, of size six or seven, and both Q 1 , Q 2 Q_1,Q_2 Q 1 , Q 2 occur in it. At each prime, skewness requires at least two positive Smith exponents of that generically injective linear subpencil (Lemma 15.2 with e = 1 e = 1 e = 1 ), so Q 1 2 Q 2 2 Q_1^2Q_2^2 Q 1 2 Q 2 2 divides every maximal minor, of degree at most seven, while Q 1 2 Q 2 2 Q_1^2Q_2^2 Q 1 2 Q 2 2 has degree eight and some maximal minor is nonzero: a contradiction.
Descent and alignment of the two components
Work over the field R ( X ) \mathbb{R}(X) R ( X ) . Let C \mathcal{C} C be a set of geometric factors of D D D , that is, irreducible factors over an algebraic closure of R ( X ) \mathbb{R}(X) R ( X ) , of total Y Y Y -degree two, that is invariant under the Galois group. Their product, normalized so that a chosen nonzero coefficient equals 1, is Galois-invariant and so has coefficients in R ( X ) \mathbb{R}(X) R ( X ) ; clearing denominators and dividing by the content in R [ X ] \mathbb{R}[X] R [ X ] gives, by Gauss’s lemma, a real polynomial factor Q ( X , Y ) Q(X,Y) Q ( X , Y ) of D D D of Y Y Y -degree two. Every irreducible factor of the bihomogeneous D D D is bihomogeneous. Under Y ↦ Y + t X Y \mapsto Y+tX Y ↦ Y + tX a factor maps to a factor of D D D with the same highest Y Y Y -degree part; being bihomogeneous, the image factor equals that part, so the factor is invariant, and likewise in the other slot. The shears generate ( X , Y ) ↦ ( t X , t − 1 Y ) (X,Y) \mapsto(tX,t^{-1}Y) ( X , Y ) ↦ ( tX , t − 1 Y ) , forcing equal degrees, and ( X , Y ) ↦ ( Y , − X ) (X,Y) \mapsto(Y,-X) ( X , Y ) ↦ ( Y , − X ) , giving symmetry in bidegree ( 2 , 2 ) (2,2) ( 2 , 2 ) ; this extends to descended products, factor by factor. At a generic real anchor Q ( X , ⋅ ) Q(X,\cdot) Q ( X , ⋅ ) is a real quadratic factor of D X D_X D X , which is Q 2 Q^2 Q 2 or Q 1 Q 2 Q_1Q_2 Q 1 Q 2 by Proposition 11.3, so Q ( X , ⋅ ) Q(X,\cdot) Q ( X , ⋅ ) is proportional to a semidefinite quadric; semidefiniteness extends to all fibres by continuity. If Q ( X 0 , ⋅ ) ≥ 0 Q(X_0,\cdot) \ge0 Q ( X 0 , ⋅ ) ≥ 0 and Q ( X 0 , Y ) > 0 Q(X_0,Y)>0 Q ( X 0 , Y ) > 0 , then Q ( Y , X 0 ) = Q ( X 0 , Y ) > 0 Q(Y,X_0)=Q(X_0,Y)>0 Q ( Y , X 0 ) = Q ( X 0 , Y ) > 0 by symmetry, so the semidefinite fibre Q ( Y , ⋅ ) Q(Y,\cdot) Q ( Y , ⋅ ) is nonnegative; such Y Y Y are dense. So Q Q Q is the sectional form of a nonnegative curvature family, which is aligned if its factors are aligned at generic anchors. As in Remark 27.2, algebraic invariants of geometric factors (ranks, degrees, spans) persist under generic specialization.
Proposition 18.4. On a two-quadratic branch, both quadrics are aligned, the intrinsic denominator D D D is irreducible over R [ X , Y ] \mathbb{R}[X,Y] R [ X , Y ] , and its geometric factors have equal algebraic invariants. The case in which both quadrics have rank eight is excluded.
Proof. By Lemma 10.4, Q 1 2 Q 2 2 Q_1^2Q_2^2 Q 1 2 Q 2 2 divides det P \det P det P , and both have degree eight, so det P = c Q 1 2 Q 2 2 \det P=cQ_1^2Q_2^2 det P = c Q 1 2 Q 2 2 . At v = 0 v=0 v = 0 the radial row vanishes, so Q 1 ( 0 , u ) Q 2 ( 0 , u ) = 0 Q_1(0,u)Q_2(0,u)=0 Q 1 ( 0 , u ) Q 2 ( 0 , u ) = 0 identically. One of the two restrictions is identically zero, and semidefiniteness puts that real subspace in its radical, so at least one quadric is aligned. If exactly one is aligned, the geometric factors vanishing on ker J X \ker\mathcal{J}_X ker J X form a rational, Galois-invariant collection of total degree two (a rank-two real semidefinite quadric contributes both conjugate linear factors or neither, since a complex linear form vanishes on a real subspace exactly when its conjugate does), and the descent gives a nonzero nonnegative aligned family, contrary to Proposition 18.3. So both are aligned.
A proper real factor of D D D would specialize to a divisor of Q 1 Q 2 Q_1Q_2 Q 1 Q 2 of degree two (factors in one slot only are excluded), and the same shear and sign argument would give an aligned family excluded by Proposition 18.3. So D D D is irreducible over R [ X , Y ] \mathbb{R}[X,Y] R [ X , Y ] , its primitive fibre is irreducible over R ( X ) [ Y ] \mathbb{R}(X)[Y] R ( X ) [ Y ] , and its geometric factors form one Galois orbit, whose degrees, ranks, support dimensions and incidence invariants are equal; by Remark 27.2 these equalities persist at generic anchors of the branch. In particular a geometric degree list 2 , 1 , 1 2,1,1 2 , 1 , 1 is impossible, and so is a geometrically irreducible quartic, since absolute irreducibility persists on a nonempty Zariski-open set of specializations meeting the branch, where D X = Q 1 Q 2 D_X=Q_1Q_2 D X = Q 1 Q 2 . This leaves two quadratics or four lines.
Finally, if both aligned quadrics have rank eight, then det P ( v , u ) = c Q 1 ( v ) 2 Q 2 ( v ) 2 ≠ 0 \det P(v,u)=cQ_1(v)^2Q_2(v)^2\ne0 det P ( v , u ) = c Q 1 ( v ) 2 Q 2 ( v ) 2 = 0 for real v ≠ 0 v\ne0 v = 0 . For a fixed nonzero real a a a the map v ↦ P ( v , 0 ) a v\mapsto P(v,0)a v ↦ P ( v , 0 ) a is an isomorphism of eight-spaces; if P ( 0 , u ) a ≠ 0 P(0,u)a\ne0 P ( 0 , u ) a = 0 , the vector v ≠ 0 v\ne0 v = 0 with P ( v , 0 ) a = − P ( 0 , u ) a P(v,0)a=-P(0,u)a P ( v , 0 ) a = − P ( 0 , u ) a gives P ( v , u ) a = 0 P(v,u)a=0 P ( v , u ) a = 0 , contradicting the determinant identity. So P ( 0 , u ) = 0 P(0,u)=0 P ( 0 , u ) = 0 , the full transpose image is E X E_X E X , and on the real open branch Lemmas 16.2 and 16.1 and Proposition 9.1(2) make R R R Clifford, contrary to our assumption. □
Quadratic poles with proper coefficient support
Retain the notation of Section 18, including sole-quadratic and two-quadratic branches. For a simple real irreducible quadratic pole Q Q Q , let U ⊂ R 8 U\subset\mathbb{R}^8 U ⊂ R 8 be the real support (Section 17) of its skew residue, d = dim U d=\dim U d = dim U . The residue has rank two: it is skew, so its rank is even; Q Q Q raised to that rank divides det P \det P det P , of degree eight (Lemma 10.4), so the rank is at most four; and rank four is excluded by Proposition 14.1. For each geometric pole component let A A A now denote the closed projective residue image, α = dim A \alpha= \dim A α = dim A , s ( a ) = rank B ( a ) s(a) = \operatorname{rank} B(a) s ( a ) = rank B ( a ) , and G ( a ) = B ( a ) B ( a ) T G(a) = B(a)B(a)^{T} G ( a ) = B ( a ) B ( a ) T . Lemmas 17.4 and 10.3 give
4 ≤ s ( a ) ≤ α ≤ 5 ( a ∈ A ) , A ⊂ V ( a T H a ) , rank G ( a ) = 2 s ( a ) − 8 , 4 \le s(a) \le\alpha\le5 \quad(a \in A), \qquad A \subset V(a^{T}Ha), \qquad\operatorname{rank} G(a) = 2s(a) - 8, 4 ≤ s ( a ) ≤ α ≤ 5 ( a ∈ A ) , A ⊂ V ( a T H a ) , rank G ( a ) = 2 s ( a ) − 8 ,
and A A A has no real points. Supports d ≤ 5 d \le5 d ≤ 5 are excluded by Lemma 17.1 ; this section treats d = 6 , 7 d = 6, 7 d = 6 , 7 , including rank-two real quadrics that split over C \mathbb{C} C .
Six-dimensional support
Proposition 19.1. If d = 6 d = 6 d = 6 , then H ∣ U H|_{U} H ∣ U is definite, and with h = ± H ∣ U h = \pm H|_{U} h = ± H ∣ U chosen positive definite one has G ( a ) = h ( a ) D G(a) = h(a)D G ( a ) = h ( a ) D on U U U with D ⪰ 0 D \succeq0 D ⪰ 0 of rank eight and image E E E . The pole quadric Q Q Q is proportional to Y T D Y Y^{T}DY Y T D Y , has rank eight and has kernel K ∩ X ⊥ \mathcal{K} \cap X^{\perp} K ∩ X ⊥ . Two distinct quadratic poles cannot both have support dimension six.
Proof. H ∣ U ≠ 0 H|_{U} \ne0 H ∣ U = 0 , since a nondegenerate eight-form has no totally isotropic six-space, so its projective null locus in P ( U C ) \mathbb{P}(U_{\mathbb{C}}) P ( U C ) has dimension four and each residue image fills a component. If the restricted rank is at least three this quadric is irreducible, and an indefinite or degenerate real restriction would give a real point; at ranks one or two every hyperplane component contains the nonzero real radical, contrary to the absence of real points in either case. So H ∣ U H|_{U} H ∣ U is definite of rank six. The image is its whole null quadric, α = s = 4 \alpha= s = 4 α = s = 4 , so G = 0 G = 0 G = 0 there, and, each entry vanishing on the irreducible quadric h = 0 h = 0 h = 0 , quadratic divisibility gives G = h D G = hD G = h D ; evaluation at a real a ≠ 0 a \ne0 a = 0 makes D D D positive semidefinite of rank eight. All B ( a ) B(a) B ( a ) , a ∈ U ∖ 0 a \in U \setminus0 a ∈ U ∖ 0 , have the image Im D \operatorname{Im} D Im D , which is E E E by Lemma 12.5 (1).
Contraction and polarization give ( P ( Y ) ∣ U ) T ( P ( Y ) ∣ U ) = ( Y T D Y ) H 0 (P(Y)|_{U})^{T}(P(Y)|_{U}) = (Y^{T}DY)H_{0} ( P ( Y ) ∣ U ) T ( P ( Y ) ∣ U ) = ( Y T D Y ) H 0 with H 0 > 0 H_{0} > 0 H 0 > 0 representing h h h . Modulo Q Q Q the restricted pencil has corank at least two, so its Gram determinant vanishes and Q ∣ ( Y T D Y ) 6 Q \mid(Y^{T}DY)^{6} Q ∣ ( Y T D Y ) 6 ; equal degrees make Q Q Q proportional to Y T D Y Y^{T}DY Y T D Y , and Im D = E \operatorname{Im} D = E Im D = E gives its rank and kernel. For two six-spaces, evaluating G = h i D i G = h_{i}D_{i} G = h i D i at a real nonzero vector of their intersection (two six-spaces in R 8 \mathbb{R}^{8} R 8 meet in dimension at least four) makes D 1 , D 2 D_{1}, D_{2} D 1 , D 2 , and so the two quadrics, proportional; proportional quadrics define the same pole.
Proposition 19.2. No sole-quadratic or two-quadratic branch has a component of support dimension six.
Proof. For a sole quadratic, Proposition 19.1 gives an aligned auxiliary family of rank eight on the branch, excluded by Proposition 18.2. For two quadrics there is at most one six-support component; it has rank eight, so it is geometrically irreducible. Its support dimension is intrinsic (multiply the frame-free cross matrix by the pole factor, reduce at its generic point and take the constant span of the image, which equals the coefficient support dimension in any eigenframe), so the Galois automorphisms permuting the geometric factors over R ( X ) \mathbb{R}(X) R ( X ) (Remark 27.2 ) preserve it and the degree. A nontrivial conjugate of this factor would be the other quadric, also of support six, which is impossible; so the factor descends over R ( X ) \mathbb{R}(X) R ( X ) , and primitive descent and the shear argument of § 18 give a nonzero nonnegative aligned family, excluded by Proposition 18.3.
Seven-dimensional support with transpose rank four
Lemma 19.3. Suppose d = 7 d = 7 d = 7 and the generic transpose rank on the residue image components is four. Then the quadratic entries of G ∣ U G|_{U} G ∣ U span at most two scalar forms, Im B ( a ) = E \operatorname{Im} B(a) = E Im B ( a ) = E for every real nonzero a ∈ U a \in U a ∈ U , and Q Q Q has rank eight and kernel K ∩ X ⊥ \mathcal{K} \cap X^{\perp} K ∩ X ⊥ .
Proof. Put h = H ∣ U h=H|_U h = H ∣ U , of rank at least six; G = 0 G=0 G = 0 on the residue image. If Q Q Q is geometrically irreducible, its image is real, irreducible and spans P ( U C ) = P 6 \mathbb{P}(U_{\mathbb{C}})=\mathbb{P}^6 P ( U C ) = P 6 . If α = 5 \alpha=5 α = 5 it is the whole h h h -quadric and every entry of G G G is a multiple of h h h . If α = 4 \alpha=4 α = 4 , its degree is at least three [10, p. 1, ( ∗ ) (*) ( ∗ ) ] and even, since a generic real complementary plane meets it in conjugate pairs; an entry f f f of G G G independent of h h h puts it in the degree-four complete intersection ( h , f ) (h,f) ( h , f ) , so its degree is four and it exhausts that intersection with its scheme structure (unmixed, generically reduced), so all entries lie in span { h , f } \operatorname{span}\{h,f\} span { h , f } . If Q Q Q has rank two and both components have the same image, that image is real-defined and spans P ( U C ) \mathbb{P}(U_{\mathbb{C}}) P ( U C ) , as before. Otherwise let d 1 d_1 d 1 be the complex span dimension of one of the two conjugate images. d 1 ≤ 5 d_1\leq5 d 1 ≤ 5 is impossible: by dimension the image would be linear, and its intersection with its conjugate would contain real vectors, that is, real points. If d 1 = 6 d_1=6 d 1 = 6 , the image is the whole irreducible quadric in its hyperplane, the conjugate hyperplanes are distinct, and their union is the reduced complete intersection ( h , ℓ ℓ ˉ ) (h,\ell\bar{\ell}) ( h , ℓ ℓ ˉ ) , so all entries lie in span { h , ℓ ℓ ˉ } \operatorname{span}\{h,\ell\bar{\ell}\} span { h , ℓ ℓ ˉ } . If d 1 = 7 d_1=7 d 1 = 7 , dimension five gives the whole h h h -quadric; in dimension four a real image gives span { h , f } \operatorname{span}\{h,f\} span { h , f } by the parity argument above, while two distinct nondegenerate conjugate images have total degree at least six, too much for V ( h , f ) V(h,f) V ( h , f ) , so then no entry is independent of h h h . In all cases the entries span at most two forms.
By Lemma 12.2 (two forms, n = 7 n=7 n = 7 ) all G ( a ) G(a) G ( a ) , a ∈ U ∖ 0 a\in U\setminus0 a ∈ U ∖ 0 , have one image, which is E E E by Lemma 12.5(1). So P ( Y ) ∣ U P(Y)|_U P ( Y ) ∣ U depends only on v v v and is injective for real v ≠ 0 v\ne0 v = 0 : B ( a ) T M X v = P ( v ) a ≠ 0 B(a)^T M_Xv=P(v)a\ne0 B ( a ) T M X v = P ( v ) a = 0 . Every maximal seven-minor is divisible by Q Q Q and some is nonzero, and it depends only on v v v , so Q Q Q does too; a nonzero real vector in a smaller-rank semidefinite Q ( v ) Q(v) Q ( v ) would contradict injectivity. Hence rank Q = 8 \operatorname{rank}Q=8 rank Q = 8 with kernel K ∩ X ⊥ \mathcal{K}\cap X^\perp K ∩ X ⊥ ; in particular Q Q Q does not have rank two.
Lemma 19.4. In the preceding situation G ∣ U = h 0 D 0 G|_U=h_0D_0 G ∣ U = h 0 D 0 for a positive definite quadratic h 0 h_0 h 0 and a fixed rank-eight D 0 ⪰ 0 D_0\succeq0 D 0 ⪰ 0 , and ( P ∣ U ) T ( P ∣ U ) = Q H 1 (P|_U)^T(P|_U)=QH_1 ( P ∣ U ) T ( P ∣ U ) = Q H 1 after a constant rescaling, with H 1 H_1 H 1 nondegenerate. Such a component cannot coexist with a distinct quadratic pole.
Proof. Identify E E E with the coefficient space R 8 \mathbb{R}^8 R 8 through the orthonormal frame M X M_X M X , and put B E ( a ) = M X T B ( a ) B_E(a)=M_X^TB(a) B E ( a ) = M X T B ( a ) and G E ( a ) = B E ( a ) B E ( a ) T G_E(a)=B_E(a)B_E(a)^T G E ( a ) = B E ( a ) B E ( a ) T ; for a ∈ U a\in U a ∈ U the image of B ( a ) B(a) B ( a ) is E E E (proof of Lemma 19.3), so G ( a ) = M X G E ( a ) M X T G(a)=M_XG_E(a)M_X^T G ( a ) = M X G E ( a ) M X T . The square matrix B E ( a ) B_E(a) B E ( a ) is invertible at real nonzero a ∈ U a\in U a ∈ U and B E ( a ) g X = − a B_E(a)g_X=-a B E ( a ) g X = − a with ∥ g X ∥ = 1 \lVert g_X\rVert=1 ∥ g X ∥ = 1 , so a T G E ( a ) − 1 a = 1 a^TG_E(a)^{-1}a=1 a T G E ( a ) − 1 a = 1 . If the scalar span is two-dimensional, the pencil G E G_E G E contains the positive definite G E ( a 0 ) G_E(a_0) G E ( a 0 ) , and simultaneous congruence gives G E ( a ) = S diag ( f j ( a ) I b j ) S T G_E(a)=S\operatorname{diag}(f_j(a)I_{b_j})S^T G E ( a ) = S diag ( f j ( a ) I b j ) S T , ∑ b j = 8 \sum b_j=8 ∑ b j = 8 , with pairwise nonproportional forms f j f_j f j of the span, each positive definite on the seven-space, hence irreducible and pairwise nonassociate. The inverse identity reads ∑ j n j ( a ) / f j ( a ) = 1 \sum_j n_j(a)/f_j(a)=1 ∑ j n j ( a ) / f j ( a ) = 1 with n j = ∥ pr j S − 1 a ∥ 2 n_j=\lVert\operatorname{pr}_jS^{-1}a\rVert^2 n j = ∥ pr j S − 1 a ∥ 2 ; clearing denominators and reducing modulo f j f_j f j gives f j ∣ n j f_j\mid n_j f j ∣ n j . A nonzero n j n_j n j has rank seven but a sum-of-squares presentation of rank at most b j b_j b j , and some n j ≠ 0 n_j\ne0 n j = 0 , so the blocks have sizes seven and one, and ( det B E ) 2 = ( det S ) 2 f 1 2 f 2 (\det B_E)^2=(\det S)^2f_1^2f_2 ( det B E ) 2 = ( det S ) 2 f 1 2 f 2 , for the forms $f_1,f_2 of the blocks of sizes seven and one, has odd valuation, a contradiction. So the span is one form, contraction gives ( P ∣ U ) T ( P ∣ U ) = ( Y T D 0 Y ) H 1 (P|_U)^T(P|_U)=(Y^TD_0Y)H_1 ( P ∣ U ) T ( P ∣ U ) = ( Y T D 0 Y ) H 1 , and its determinant modulo Q Q Q makes Q Q Q proportional to Y T D 0 Y Y^TD_0Y Y T D 0 Y . Modulo Q Q Q the image of the eight-by-seven P ∣ U P|_U P ∣ U is totally isotropic in the standard eight-space, of dimension at most four; with the last column, rank P ≤ 5 \operatorname{rank}P\leq5 rank P ≤ 5 there, so Q 3 ∣ det P Q^3\mid\det P Q 3 ∣ det P . A distinct quadratic pole Q ′ Q' Q ′ needs ( Q ′ ) 2 ∣ det P (Q')^2\mid\det P ( Q ′ ) 2 ∣ det P (Lemma 10.4), and Q 3 ( Q ′ ) 2 Q^3(Q')^2 Q 3 ( Q ′ ) 2 has degree ten > > > deg det P = 8 \deg\det P=8 deg det P = 8 .
Proposition 19.5. A two-quadratic branch has no component of support dimension seven.
Proof. Write det P = c Q 1 2 Q 2 2 \det P=cQ_1^2Q_2^2 det P = c Q 1 2 Q 2 2 . The Smith exponents at either prime are ( 1 , 1 , 0 6 ) (1,1,0^6) ( 1 , 1 , 0 6 ) , so every seven-minor is divisible by both and P ♯ = adj P / ( c Q 1 Q 2 ) P^\sharp=\operatorname{adj}P/(cQ_1Q_2) P ♯ = adj P / ( c Q 1 Q 2 ) is cubic with P ♯ P = P P ♯ = Q 1 Q 2 I P^\sharp P=PP^\sharp=Q_1Q_2I P ♯ P = P P ♯ = Q 1 Q 2 I . If the first residue lies in U = ker f U=\ker f U = ker f , then f P ♯ = Q 1 ℓ ( Y ) T fP^\sharp=Q_1\ell(Y)^T f P ♯ = Q 1 ℓ ( Y ) T with ℓ \ell ℓ linear, since the image of P ♯ P^\sharp P ♯ modulo Q 1 Q_1 Q 1 is the residue kernel plane; multiplying by P P P gives P ( Y ) T ℓ ( Y ) = Q 2 ( Y ) f T P(Y)^T\ell(Y)=Q_2(Y)f^T P ( Y ) T ℓ ( Y ) = Q 2 ( Y ) f T . Let Θ : Y ↦ ℓ ( Y ) \Theta:Y\mapsto\ell(Y) Θ : Y ↦ ℓ ( Y ) and, for real nonzero a ∈ U a\in U a ∈ U , P a Y = P ( Y ) a P_aY=P(Y)a P a Y = P ( Y ) a . By the last identity, as f a = 0 f_a=0 f a = 0 , P a T Θ P_a^T\Theta P a T Θ is skew, and surjectivity of P a P_a P a gives it kernel ker Θ \ker\Theta ker Θ and even rank m = rank Θ m=\operatorname{rank}\Theta m = rank Θ , with the same image for every such a a a . For nonzero real g ∈ Im Θ g\in\operatorname{Im}\Theta g ∈ Im Θ the vectors N g a N_g a N g a , a ∈ U a\in U a ∈ U , lie in that image, and N g N_g N g is injective, so m ≥ 7 m\ge7 m ≥ 7 , hence m = 8 m=8 m = 8 and Θ \Theta Θ is surjective. Over C \mathbb{C} C , rank ( P a T Θ ) = s ( a ) \operatorname{rank}(P_a^T\Theta)=s(a) rank ( P a T Θ ) = s ( a ) is then even for every a ∈ U C a\in U_{\mathbb{C}} a ∈ U C , so, as 4 ≤ s ≤ 5 4\le s\le5 4 ≤ s ≤ 5 (Lemma 17.4), s = 4 s=4 s = 4 on each residue image, and Lemma 19.4 gives the degree contradiction. This covers conjugate geometric factors of a rank-two quadric as well.
Seven-dimensional support with transpose rank five
Here α = s = 5 \alpha=s=5 α = s = 5 . As U U U is a hyperplane, H ∣ U H|_U H ∣ U has rank at least six and its null quadric in P 6 \mathbb{P}^6 P 6 is irreducible and filled by the residue image; absence of real points makes H ∣ U H|_U H ∣ U definite of rank seven. Reversing the curvature sign if necessary, h = H ∣ U > 0 h=H|_U>0 h = H ∣ U > 0 and H H H has signature ( 7 , 1 ) (7,1) ( 7 , 1 ) .
Lemma 19.6. In this rank-five case, either Im B ( a ) = E \operatorname{Im}B(a)=E Im B ( a ) = E for all real a ∈ U ∖ 0 a\in U\setminus0 a ∈ U ∖ 0 , or
G = h D + u w T − v v T (50) G=hD+uw^T-vv^T \tag*{(50)} G = h D + u w T − v v T (50) on U U U , with D D D constant symmetric and u , v u,v u , v real linear columns.
Proof. The cone ring S = C [ U ] / ( h ) S=\mathbb{C}[U]/(h) S = C [ U ] / ( h ) is factorial by the Klein–Nagata theorem [29 Theorem 8.2, p. 26] , h h h being nondegenerate in seven variables. By Lemma 12.7 the induced pairing of G G G is hyperbolic over the real function field, and the saturated isotropic line modules are rank-one reflexive, hence principal. So Lemma 12.8 applies over R [ U ] / ( h ) \mathbb{R}[U]/(h) R [ U ] / ( h ) ; lifting its columns to real polynomials gives (19.1) or G = h D + c b T + b c T G=hD+cb^T+bc^T G = h D + c b T + b c T with c ≠ 0 c\ne0 c = 0 constant and b b b quadratic. In the second case, compression to c ⊥ c^\perp c ⊥ is h D 0 hD_0 h D 0 , and positivity puts the cross block in Im D 0 \operatorname{Im}D_0 Im D 0 , so all images lie in a fixed space of dimension at most nine and Lemma 12.5 (2) gives the first alternative.
Lemma 19.7. If the first alternative of Lemma 19.6 fails, then D ⪰ 0 D\succeq0 D ⪰ 0 in (19.1) has rank exactly eight. A constant hyperbolic change of the two columns gives
G = h D − w w T + z z T , w ∈ F : = Im D , w T D + w = h (51) G=hD-ww^T+zz^T,\qquad w\in\mathcal{F}:=\operatorname{Im}D,\qquad w^TD^+w=h \tag*{(51)} G = h D − w w T + z z T , w ∈ F := Im D , w T D + w = h (51) where the projection of z z z to F ⊥ \mathcal{F}^{\perp} F ⊥ has rank at least two.
Proof. For a fixed real vector b ∈ V b\in V b ∈ V choose nonzero a ∈ U a\in U a ∈ U with b T u ( a ) = b T v ( a ) = 0 b^Tu(a)=b^Tv(a)=0 b T u ( a ) = b T v ( a ) = 0 ; positivity gives b T D b ≥ 0 b^TDb\ge0 b T D b ≥ 0 . Put r = rank D r=\operatorname{rank}D r = rank D and G + = h D + u u T G_+=hD+uu^T G + = h D + u u T . Positivity of G = G + − v v T G=G_+-vv^T G = G + − v v T gives v ∈ Im G + v\in\operatorname{Im}G_+ v ∈ Im G + , and rank eight gives rank G + ∈ { 8 , 9 } \operatorname{rank}G_+\in\{8,9\} rank G + ∈ { 8 , 9 } , so 7 ≤ r ≤ 9 7\le r\le9 7 ≤ r ≤ 9 . If r = 9 r=9 r = 9 , u , v u,v u , v lie in F \mathcal{F} F and Lemma 12.5 (2) aligns the image.
Suppose r = 7 r=7 r = 7 . The quotient map u ˉ : U → V / F \bar{u}:U\to V/\mathcal{F} u ˉ : U → V / F is injective and Im B ( a ) = F + R u ( a ) \operatorname{Im}B(a)=\mathcal{F}+\mathbb{R}u(a) Im B ( a ) = F + R u ( a ) for real a ≠ 0 a\ne0 a = 0 , a space containing the radial vector M X a M_Xa M X a . Two pointwise proportional linear maps, one of rank at least two, have constant ratio (compare two independent images and their sum), so M X a = c u ˉ ( a ) M_Xa=c\bar{u}(a) M X a = c u ˉ ( a ) for a constant c c c . If c = 0 c=0 c = 0 , then F = M X U ⊂ E \mathcal{F}=M_XU\subset E F = M X U ⊂ E , each lower image has dimension at most one and, being skew, it is zero; then the injective u ˉ ( U ) \bar{u}(U) u ˉ ( U ) would lie in the one-dimensional E / F E/\mathcal{F} E / F , which is impossible. If c ≠ 0 c\ne0 c = 0 , the lower skew pencil has one fixed image K X T F K_X^T\mathcal{F} K X T F at all nonzero real coefficients, of even rank 0 , 2 , 4 0,2,4 0 , 2 , 4 or 6 6 6 . Ranks two and four are impossible, as their skew-operator spaces have dimensions one and six, less than seven; rank six would give a cubic Pfaffian, an odd function, nowhere zero on the connected real unit sphere, which is impossible; rank zero puts F ⊂ E \mathcal{F}\subset E F ⊂ E , contradicting the rank of the quotient radial map. So r ≠ 7 r\ne7 r = 7 .
Thus r = 8 r=8 r = 8 . If rank u ˉ ≤ 1 \bar{u}\le1 u ˉ ≤ 1 , all images lie in a fixed nine-space and Lemma 12.5 (2) aligns them. Otherwise positivity gives v ˉ = c ˉ u ˉ \bar{v}=\bar{c}\,\bar{u} v ˉ = c ˉ u ˉ with a constant c c c , so v = c u + w 0 v=cu+w_0 v = c u + w 0 with w 0 ∈ F w_0\in\mathcal{F} w 0 ∈ F , and where u ˉ ≠ 0 \bar{u}\ne0 u ˉ = 0 the rank drop from nine to eight reads c 2 + h − 1 w 0 T D + w 0 = 1 c^2+h^{-1}w_0^TD^+w_0=1 c 2 + h − 1 w 0 T D + w 0 = 1 : write G + = M + M + T G_+=M_+M_+^T G + = M + M + T with M + = [ h S , u ] M_{+}=[\sqrt{h}S,u] M + = [ h S , u ] , where S S T = D SS^{T}=D S S T = D and S S S maps R 8 \mathbb{R}^{8} R 8 isomorphically onto F \mathcal{F} F ; M + M_{+} M + is injective since u ∉ F u\notin\mathcal{F} u ∈ / F , v = M + ξ v=M_{+}\xi v = M + ξ with ξ = ( S − 1 w 0 / h , c ) \xi=(S^{-1}w_{0}/\sqrt{h},c) ξ = ( S − 1 w 0 / h , c ) , and G + − v v T = M + ( I − ξ ξ T ) M + T G_{+}-vv^{T}=M_{+}(I-\xi\xi^{T})M_{+}^{T} G + − v v T = M + ( I − ξ ξ T ) M + T has rank eight exactly when ∥ ξ ∥ 2 = 1 \lVert\xi\rVert^{2}=1 ∥ ξ ∥ 2 = 1 . Multiplied by h h h , this is a polynomial identity valid everywhere. As h > 0 h>0 h > 0 on U ∖ 0 U\setminus0 U ∖ 0 and D + ⪰ 0 D^{+}\succeq0 D + ⪰ 0 on F \mathcal{F} F , c 2 ≤ 1 c^{2}\leq1 c 2 ≤ 1 . If ∣ c ∣ = 1 \lvert c\rvert=1 ∣ c ∣ = 1 , then w 0 = 0 w_{0}=0 w 0 = 0 and the image is fixed and aligned. Otherwise put δ = 1 − c 2 > 0 \delta=1-c^{2}>0 δ = 1 − c 2 > 0 , w = w 0 / δ w=w_{0}/\sqrt{\delta} w = w 0 / δ , z = δ u − c w z=\sqrt{\delta}\,u-cw z = δ u − c w ; expanding gives (19.2), and the quotient column keeps rank at least two.
Proposition 19.8. If the first alternative of Lemma 19.6 fails on a branch with d = 7 d=7 d = 7 and generic transpose rank five, then there is a positive Jacobi eigenvalue of multiplicity at least five.
Proof. Project B B B to F ⊥ \mathcal{F}^{\perp} F ⊥ in (19.2), writing the projection as B ⊥ B_{\perp} B ⊥ and that of z z z as b b b ; then B ⊥ B ⊥ T = b b T B_{\perp}B_{\perp}^{T}=bb^{T} B ⊥ B ⊥ T = b b T . Every row of B ⊥ B_{\perp} B ⊥ has squared norm b i 2 b_i^{2} b i 2 , so it vanishes on the real hyperplane b i = 0 b_i=0 b i = 0 and is b i e i T b_i e_i^{T} b i e i T for a constant unit vector e i e_i e i ; the cross-row identities give e i T e j = 1 e_i^{T}e_j=1 e i T e j = 1 , so all e i e_i e i equal one unit vector e e e and B ⊥ = b e T B_{\perp}=be^{T} B ⊥ = b e T . The mixed block gives B ( a ) e = z ( a ) B(a)e=z(a) B ( a ) e = z ( a ) . In orthonormal input coordinates beginning with e e e , B = [ z , B 0 ] B=[z,B_0] B = [ z , B 0 ] with B 0 B_0 B 0 valued in F \mathcal{F} F and B 0 B 0 T = h D − w w T B_0B_0^{T}=hD-ww^{T} B 0 B 0 T = h D − w w T . Choose S : R 8 → F S:\mathbb{R}^{8}\to\mathcal{F} S : R 8 → F with S S T = D SS^{T}=D S S T = D , put Ξ 0 = S − 1 B 0 \Xi_0=S^{-1}B_0 Ξ 0 = S − 1 B 0 and η 0 = S − 1 w \eta_0=S^{-1}w η 0 = S − 1 w ; then ∥ η 0 ∥ 2 = w T D + w = h \lVert\eta_0\rVert^{2}=w^{T}D^{+}w=h ∥ η 0 ∥ 2 = w T D + w = h by (19.2), and the seven nonzero eigenvalues of Ξ 0 Ξ 0 T = h I − η 0 η 0 T \Xi_0\Xi_0^{T}=hI-\eta_0\eta_0^{T} Ξ 0 Ξ 0 T = h I − η 0 η 0 T are h h h , so Ξ 0 T Ξ 0 = h I 7 \Xi_0^{T}\Xi_0=hI_7 Ξ 0 T Ξ 0 = h I 7 .
For g ∈ e ⊥ g\in e^{\perp} g ∈ e ⊥ , using B ( a ) g = N g a B(a)g=N_ga B ( a ) g = N g a and an h h h -orthonormal basis of U U U , the eight-by-seven matrix Ξ ( g ) \Xi(g) Ξ ( g ) , linear in the seven variables g g g , satisfies Ξ ( g ) T Ξ ( g ) = q ( g ) I 7 \Xi(g)^{T}\Xi(g)=q(g)I_7 Ξ ( g ) T Ξ ( g ) = q ( g ) I 7 with q ( g ) = ∥ g ∥ 2 q(g)=\lVert g\rVert^{2} q ( g ) = ∥ g ∥ 2 . At the prime ( q ) (q) ( q ) its image is isotropic in the standard eight-space, of dimension at most four, so each maximal seven-minor is divisible by q 3 q^{3} q 3 ; the signed cofactor vector is q 3 η ( g ) q^{3}\eta(g) q 3 η ( g ) with η \eta η linear, Cauchy–Binet gives ∥ η ∥ 2 = q \lVert\eta\rVert^{2}=q ∥ η ∥ 2 = q , and the normal-vector identity gives Ξ ( g ) Ξ ( g ) T = q ( g ) I 8 − η ( g ) η ( g ) T \Xi(g)\Xi(g)^{T}=q(g)I_8-\eta(g)\eta(g)^{T} Ξ ( g ) Ξ ( g ) T = q ( g ) I 8 − η ( g ) η ( g ) T .
Complete the h h h -orthonormal basis of U U U by its H H H -orthogonal negative unit vector; in these coefficient coordinates H = Λ = diag ( I 7 , − 1 ) H=\Lambda=\operatorname{diag}(I_7,-1) H = Λ = diag ( I 7 , − 1 ) , the reciprocal congruence on the factors leaving the Jacobi family unchanged. With n ( g ) n(g) n ( g ) the last column of N g N_g N g in this basis and p ( g ) = S η ( g ) p(g)=S\eta(g) p ( g ) = S η ( g ) ,
J K X g = q ( g ) D − p ( g ) p ( g ) T − n ( g ) n ( g ) T , p ( g ) T D + p ( g ) = q ( g ) . (52) \mathcal{J}_{KX_g}=q(g)D-p(g)p(g)^{T}-n(g)n(g)^{T},\qquad p(g)^{T}D^{+}p(g)=q(g). \tag*{(52)} J K X g = q ( g ) D − p ( g ) p ( g ) T − n ( g ) n ( g ) T , p ( g ) T D + p ( g ) = q ( g ) . (52)
In Frac ( R [ g ] / ( q ) ) \operatorname{Frac}(\mathbb{R}[g]/(q)) Frac ( R [ g ] / ( q )) the left side is nilpotent, and the determinant identity makes the symmetric 2 × 2 2\times2 2 × 2 matrix [ p , n ] T [ p , n ] [p,n]^{T}[p,n] [ p , n ] T [ p , n ] nilpotent, of the form [ a b b − a ] \begin{bmatrix}a&b\\b&-a\end{bmatrix} [ a b b − a ] with a 2 + b 2 = 0 a^{2}+b^{2}=0 a 2 + b 2 = 0 . The rank-seven quadric is geometrically integral over R \mathbb{R} R , so its function field has no square root of − 1 -1 − 1 ; hence [ p , n ] T [ p , n ] = 0 [p,n]^{T}[p,n]=0 [ p , n ] T [ p , n ] = 0 there, and quadratic divisibility gives p T p = β q p^{T}p=\beta q p T p = β q with β > 0 \beta>0 β > 0 .
The image W 0 = Im p W_{0}=\operatorname{Im}p W 0 = Im p is a seven-space in F \mathcal{F} F , and the two norm identities in (19.3) give D + ∣ W 0 = β − 1 I D^{+}|_{W_{0}}=\beta^{-1}I D + ∣ W 0 = β − 1 I as a quadratic compression. A symmetric matrix with scalar compression on a hyperplane has that eigenvalue on the intersection of the hyperplane with the orthogonal complement of its cross-block vector, so D D D has eigenvalue β \beta β on a space L β L_{\beta} L β of dimension at least six. If this dimension is at least seven, subtracting the two rank-one terms leaves multiplicity at least five. If it is six, choose nonzero g g g in the kernel of g ↦ p L β p ( g ) g\mapsto p_{L_{\beta}}p(g) g ↦ p L β p ( g ) , from seven variables to six; the vectors of L β L_{\beta} L β orthogonal to n ( g ) n(g) n ( g ) give a β q ( g ) \beta q(g) β q ( g ) -eigenspace of dimension at least five. Constancy of the real Osserman spectrum proves the claim.
Proposition 19.9. No sole-quadratic branch whose simple pole has a rank-two residue has support dimension seven.
Proof. For generic transpose rank four, Lemma 19.3 makes the pencil aligned on U U U and the pole of rank eight, which Proposition 18.2 excludes. For rank five, Lemma 19.6 gives either the same alignment, whose pole has rank eight by the maximal-minor argument of Lemma 19.3 (which uses alignment only), or the form (19.2), with signature ( 7 , 1 ) (7,1) ( 7 , 1 ) , where Proposition 19.8 forces a positive eigenvalue of multiplicity at least five. Among the eleven patterns listed at the start of Section 18, only ( 7 , 5 , 3 ) (7,5,3) ( 7 , 5 , 3 ) gives Λ \Lambda Λ an eigenvalue of multiplicity five, and then Λ \Lambda Λ has no simple eigenvalue, so H = Λ − 1 H=\Lambda^{-1} H = Λ − 1 cannot have signature ( 7 , 1 ) (7,1) ( 7 , 1 ) up to sign; in every other pattern each eigenvalue of Λ \Lambda Λ has multiplicity at most four: a contradiction in every case.
Four-dimensional residue images
We work in Setting (T) of § 12 , with ν = 8 \nu=8 ν = 8 and Λ \Lambda Λ indefinite, and let h ( a ) = a T H a h(a)=a^{T}Ha h ( a ) = a T H a , H = Λ − 1 H=\Lambda^{-1} H = Λ − 1 , be the nondegenerate indefinite coefficient form. So G ( a ) G(a) G ( a ) is positive semidefinite of rank eight at every nonzero real a a a . Let H \mathcal{H} H be the real span of h h h and the entries G i j G_{ij} G ij , and let V ( h , G ) ⊂ P 7 V(h,G)\subset\mathbb{P}^{7} V ( h , G ) ⊂ P 7 be their common zero locus; it has no real point, since G ( a ) G(a) G ( a ) has rank eight at every nonzero real a a a . A maximal isotropic subspace of the complex form h h h has dimension four, so V ( h ) V(h) V ( h ) contains no linear P 4 \mathbb{P}^{4} P 4 . Complex varieties carry their reduced structure unless a scheme is specified. By Lemma 12.2 and Remark 12.3 (the scalar-span criterion), if the real span of the quadratic entries of G G G has dimension at most three, then
Im B ( a ) = E ( a ∈ R 8 ∖ { 0 } ) . (53) \operatorname{Im} B(a)=E \qquad(a\in\mathbb{R}^{8}\setminus\{0\}). \tag*{(53)} Im B ( a ) = E ( a ∈ R 8 ∖ { 0 }) . (53)
The quadratic base
Lemma 20.1. Let A 1 , … , A k ⊂ V ( h , G ) A_1,\ldots,A_k\subset V(h,G) A 1 , … , A k ⊂ V ( h , G ) be irreducible fourfolds whose union spans P 7 \mathbb{P}^{7} P 7 . Then either (20.1) holds, or the ideal generated by H \mathcal{H} H has height three and H \mathcal{H} H contains a real regular sequence of three quadrics whose degree-eight complete intersection contains every A i A_i A i .
Proof. The locus Z = V ( h , G ) Z=V(h,G) Z = V ( h , G ) contains the A i A_i A i , so it has dimension at least four. A six-dimensional component would be the irreducible quadric V ( h ) V(h) V ( h ) , which has real points. Some real entry f f f of G G G is independent of h h h : otherwise every entry is a multiple of h h h , against rank eight at a nonzero real h h h -null vector. Let Y Y Y be a five-dimensional component. It is not a linear P 5 \mathbb{P}^{5} P 5 , since V ( h ) V(h) V ( h ) contains no linear P 4 \mathbb{P}^{4} P 4 , so it spans either P 7 \mathbb{P}^{7} P 7 or a hyperplane.
If Y Y Y spans P 7 \mathbb{P}^{7} P 7 , then deg Y ≥ 3 \deg Y\geq3 deg Y ≥ 3 by the nondegenerate degree bound [10 p. 1, (*)] , and Y Y Y lies in the complete intersection V ( h , f ) V(h,f) V ( h , f ) , of pure dimension five and degree four. If Y ≠ Y ‾ Y\ne\overline{Y} Y = Y , the two conjugate components would have total degree at least six, more than deg V ( h , f ) = 4 \deg V(h,f)=4 deg V ( h , f ) = 4 ; so Y Y Y is real-defined; a generic transverse real complementary plane meets it only in conjugate nonreal pairs, so deg Y \deg Y deg Y is even, hence four. Thus Y Y Y exhausts V ( h , f ) V(h,f) V ( h , f ) with multiplicity one. A complete intersection has no embedded associated primes, so generic reducedness makes V ( h , f ) V(h,f) V ( h , f ) the reduced scheme Y Y Y ; its homogeneous ideal, a complete-intersection ideal of positive dimension, is saturated, so every quadratic entry of G G G lies in span R { h , f } \operatorname{span}_{\mathbb{R}}\{h,f\} span R { h , f } , and the scalar-span criterion gives (20.1).
If Y Y Y spans a hyperplane ℓ = 0 \ell=0 ℓ = 0 , the restriction of h h h to it has rank at least six, so Y = V ( h , ℓ ) Y=V(h,\ell) Y = V ( h , ℓ ) is an irreducible quadric and every quadratic entry of G G G is g = c g h + ℓ m g g=c_g h+\ell m_g g = c g h + ℓ m g with m g m_g m g linear. Since the union of the A i A_i A i spans P 7 \mathbb{P}^{7} P 7 , some A j A_j A j is not contained in ℓ = 0 \ell=0 ℓ = 0 ; as g g g and h h h vanish on A j A_j A j , so does ℓ m g \ell m_g ℓ m g , and integrality gives m g ∣ A j = 0 m_g|_{A_j}=0 m g ∣ A j = 0 for every g g g . The vector span of A j A_j A j has dimension at least six, since V ( h ) V(h) V ( h ) contains no linear P 4 \mathbb{P}^{4} P 4 , so the linear forms vanishing on A j A_j A j span at most two dimensions. All entries g = c g h + ℓ m g g=c_g h+\ell m_g g = c g h + ℓ m g therefore lie in one complex space of quadrics of dimension at most three; real linearly independent vectors stay complex linearly independent, so their real span has the same bound, and the scalar-span criterion gives (20.1).
In the remaining case dim Z = 4 \dim Z=4 dim Z = 4 , so the ideal generated by H \mathcal{H} H has height three. Begin with h h h and a real element of H \mathcal{H} H independent of it, and choose a third real element of H \mathcal{H} H outside the finitely many height-two associated primes of their complete intersection. None of these primes contains H \mathcal{H} H , and over the infinite field R \mathbb{R} R a finite union of proper linear subspaces cannot cover H \mathcal{H} H . This gives the regular sequence; its degree-eight complete intersection contains every A i A_i A i . □ \square □
Proposition 20.2. Under the hypotheses of Lemma 20.1 , assume the A i A_i A i are distinct and ∑ i deg A i ≥ 8 \sum_i \deg A_i \ge8 ∑ i deg A i ≥ 8 . Then (53) holds.
Proof. Otherwise the degree-eight complete intersection of Lemma 20.1 contains the distinct A i A_i A i , so the degree sum is eight and the A i A_i A i are exactly its components, each of multiplicity one. Unmixedness and generic reducedness make it their reduced union, and its homogeneous ideal, a complete-intersection ideal of positive dimension, is saturated. Every entry of G G G therefore lies in its degree-two part, spanned by the three defining quadrics, and the scalar-span criterion applies.
In particular, the proposition applies to two distinct fourfolds in V ( h , G ) V(h,G) V ( h , G ) that each span P 7 \mathbb{P}^{7} P 7 (each then has degree at least four), and to four distinct fourfolds in V ( h , G ) V(h,G) V ( h , G ) whose union spans P 7 \mathbb{P}^{7} P 7 (each has degree at least two, since V ( h ) V(h) V ( h ) contains no linear P 4 \mathbb{P}^{4} P 4 ). In the two-quadratic pole case these are the images of the two geometric quadratics and of the four geometric lines; Proposition 18.4 supplies these alternatives, Lemma 17.6 the distinctness of the images, and Remark 20.9 the vanishing of G G G on them and their combined span.
A real integral image
Proposition 20.3. Suppose that a geometrically integral real-defined fourfold A ⊂ P C 7 A \subset\mathbb{P}^{7}_{\mathbb{C}} A ⊂ P C 7 spans P C 7 \mathbb{P}^{7}_{\mathbb{C}} P C 7 and is contained in V ( h , G ) V(h,G) V ( h , G ) . Then (53) holds.
Lemma 20.4. Under the hypotheses of Proposition 20.3 , either (53) holds, or the ideal generated by H H H has height three and deg A ∈ { 4 , 6 } \deg A \in\{4,6\} deg A ∈ { 4 , 6 } .
Proof. Apply Lemma 20.1 with k = 1 k=1 k = 1 . In its second alternative, A A A is nondegenerate of codimension three, so deg A ≥ 4 \deg A \ge4 deg A ≥ 4 . As A A A is real-defined and has no real point, a generic transverse real complementary plane meets it only in conjugate nonreal pairs, so deg A \deg A deg A is even. The degree is at most eight, since A A A lies in the degree-eight complete intersection of that lemma, and deg A = 8 \deg A = 8 deg A = 8 gives (53) by Proposition 20.2 with k = 1 k=1 k = 1 .
Lemma 20.5. If deg A = 4 \deg A = 4 deg A = 4 , then (53) holds.
Proof. The variety A A A has minimal degree, so by the Del Pezzo–Bertini classification [10 Theorem 0.1] it is a rational normal scroll or a cone over the Veronese surface. A singular variety in this list has a nonempty intrinsic linear vertex, which, since A A A is real-defined, would be real-defined and contain a real point; but A A A has no real point. So A A A is smooth, its four positive scroll integers sum to deg A = 4 \deg A = 4 deg A = 4 , so all are one, and A A A is the Segre variety P 1 × P 3 ⊂ P ( C 2 ⊗ C 4 ) \mathbb{P}^{1} \times\mathbb{P}^{3} \subset\mathbb{P}(\mathbb{C}^{2} \otimes\mathbb{C}^{4}) P 1 × P 3 ⊂ P ( C 2 ⊗ C 4 ) .
Conjugation preserves the two rulings, which have different dimensions, so it lifts, up to a scalar absorbed into one factor, to σ 1 ⊗ σ 2 \sigma_1 \otimes\sigma_2 σ 1 ⊗ σ 2 with conjugate-linear maps on the two factors. Then ( σ 1 ⊗ σ 2 ) 2 = I (\sigma_1 \otimes\sigma_2)^2 = I ( σ 1 ⊗ σ 2 ) 2 = I forces σ 1 2 = α I \sigma_1^2 = \alpha I σ 1 2 = α I and σ 2 2 = β I \sigma_2^2 = \beta I σ 2 2 = β I with α β = 1 \alpha\beta= 1 α β = 1 . Each map commutes with its square, so these scalars are real; rescaling makes them signs, and they agree because the ambient real structure has square one. Positive signs would give fixed vectors in both factors, hence a real Segre point; so both signs are negative.
The standard S U ( 2 ) \mathrm{SU}(2) SU ( 2 ) -action on the first factor commutes with its quaternionic real structure σ 1 \sigma_1 σ 1 , so it acts on the real coefficient eight-space, and its three quaternion units make that space a left H \mathbb{H} H -module. The quadratic ideal of the Segre embedding, the kernel of restricting quadratic polynomials to rank-one tensors, is ⋀ 2 ( C 2 ) ∗ ⊗ ⋀ 2 ( C 4 ) ∗ \bigwedge^{2}(\mathbb{C}^{2})^{*} \otimes\bigwedge^{2}(\mathbb{C}^{4})^{*} ⋀ 2 ( C 2 ) ∗ ⊗ ⋀ 2 ( C 4 ) ∗ , on which the first S L ( 2 , C ) \mathrm{SL}(2,\mathbb{C}) SL ( 2 , C ) acts trivially. So every entry of G G G is S U ( 2 ) \mathrm{SU}(2) SU ( 2 ) -invariant, and by homogeneity G ( b a ) = ∣ b ∣ 2 G ( a ) G(ba) = |b|^{2}G(a) G ( ba ) = ∣ b ∣ 2 G ( a ) for b ∈ H b \in\mathbb{H} b ∈ H . For every nonzero real a a a , the space U = H a U = \mathbb{H}a U = H a has real dimension four, since nonzero quaternions act invertibly, and this identity makes Im G \operatorname{Im} G Im G , hence Im B \operatorname{Im} B Im B , one fixed eight-space on U ∖ { 0 } U \setminus\{0\} U ∖ { 0 } ; Lemma 12.5 (1) identifies it with E E E . Since a a a was arbitrary, (53) holds.
:::
Lemma 20.6. If deg A = 6 \deg A = 6 deg A = 6 and the ideal generated by H \mathcal{H} H has height three, there is a reduced real complete intersection C = A ∪ Q C = A \cup Q C = A ∪ Q of three quadrics, where Q Q Q is a real-defined quadric fourfold spanning P ( U C ) \mathbb{P}(U_C) P ( U C ) for a real six-space U U U on which h h h is definite.
Proof. The quadratic form t = tr G ∈ H t = \operatorname{tr} G \in\mathcal{H} t = tr G ∈ H is positive definite, so it is independent of the indefinite form h h h . By height three, prime avoidance as in Lemma 20.1 gives a real f ∈ H f \in\mathcal{H} f ∈ H such that ( h , t , f ) (h,t,f) ( h , t , f ) is a regular sequence. Its complete intersection C C C has degree eight and no real point. A four-dimensional component of degree one would be a linear P 4 \mathbb{P}^{4} P 4 in V ( h ) V(h) V ( h ) , so every component has degree at least two. The component A A A takes degree six with multiplicity one, so the rest is exactly one component Q Q Q of degree two and multiplicity one, preserved by conjugation as the unique residual component. Unmixedness and generic reducedness make C = A ∪ Q C = A \cup Q C = A ∪ Q reduced.
A fourfold of degree two spans at most a projective five-space, and Q Q Q is not linear, so it spans exactly a projective five-space P ( U C ) \mathbb{P}(U_C) P ( U C ) , for a real six-space U U U since Q Q Q is conjugation-invariant. Its equation is proportional to h ∣ U h|_U h ∣ U , which is nonzero because V ( h ) V(h) V ( h ) contains no linear P 5 \mathbb{P}^{5} P 5 . Since Q Q Q has no real point, this real form is definite.
Lemma 20.7. In Lemma 20.6 , choose real independent linear forms u , v u,v u , v with U = ker u ∩ ker v U = \ker u \cap\ker v U = ker u ∩ ker v . There are real quadrics s , r , Δ s,r,\Delta s , r , Δ such that
I C = ( h , s , r ) , ( I A ) 2 ⊂ span R { h , s , r , Δ } , s ∣ U = r ∣ U = 0 , I_C = (h,s,r), \qquad(I_A)_2 \subset\operatorname{span}_{\mathbb{R}}\{h,s,r,\Delta\}, \qquad s|_U = r|_U = 0, I C = ( h , s , r ) , ( I A ) 2 ⊂ span R { h , s , r , Δ } , s ∣ U = r ∣ U = 0 , where ( I A ) 2 (I_A)_2 ( I A ) 2 is the real degree-two part of the homogeneous ideal.
Proof. Since I Q = ( h , u , v ) I_Q = (h,u,v) I Q = ( h , u , v ) , subtracting constant multiples of h h h from t , f t,f t , f gives
s = u a 0 + v b 0 , r = u c 0 + v d 0 , Δ = a 0 d 0 − b 0 c 0 , s = ua_0 + vb_0, \qquad r = uc_0 + vd_0, \qquad\Delta= a_0d_0 - b_0c_0, s = u a 0 + v b 0 , r = u c 0 + v d 0 , Δ = a 0 d 0 − b 0 c 0 ,
with real linear forms a 0 , b 0 , c 0 , d 0 a_0,b_0,c_0,d_0 a 0 , b 0 , c 0 , d 0 . Work in the graded domain S = R [ x 1 , … , x 8 ] / ( h ) S = \mathbb{R}[x_1,\ldots,x_8]/(h) S = R [ x 1 , … , x 8 ] / ( h ) in coefficient coordinates x i x_i x i . Both ( s , r ) (s,r) ( s , r ) and ( u , v ) (u,v) ( u , v ) are regular sequences in S S S ; the latter follows from h ∣ U ≠ 0 h|_U \ne0 h ∣ U = 0 and the Cohen–Macaulay property of the polynomial ring. For g ∈ ( I A ) 2 g \in(I_A)_2 g ∈ ( I A ) 2 , reducedness gives u g , v g ∈ I A ∩ I Q = I C ug,vg \in I_A \cap I_Q = I_C ug , v g ∈ I A ∩ I Q = I C , so modulo h h h
u g = a 1 s + b 1 r , v g = c 1 s + d 1 r ug = a_1s + b_1r, \qquad vg = c_1s + d_1r ug = a 1 s + b 1 r , v g = c 1 s + d 1 r
with real linear coefficients. The Koszul syzygies of ( s , r ) (s,r) ( s , r ) , with degrees taken into account, give a real constant k k k with
v a 1 − u c 1 = k r , v b 1 − u d 1 = − k s . va_1 - uc_1 = kr, \qquad vb_1 - ud_1 = -ks. v a 1 − u c 1 = k r , v b 1 − u d 1 = − k s .
Substituting the expressions for s , r s,r s , r and using the Koszul syzygies of ( u , v ) (u,v) ( u , v ) gives real constants e 1 , e 2 e_1,e_2 e 1 , e 2 with
a 1 − k d 0 = e 1 u , c 1 + k c 0 = e 1 v , b 1 + k b 0 = e 2 u , d 1 − k a 0 = e 2 v . a_1 - kd_0 = e_1u, \qquad c_1 + kc_0 = e_1v, \qquad b_1 + kb_0 = e_2u, \qquad d_1 - ka_0 = e_2v. a 1 − k d 0 = e 1 u , c 1 + k c 0 = e 1 v , b 1 + k b 0 = e 2 u , d 1 − k a 0 = e 2 v .
Hence u g = k ( d 0 s − b 0 r ) + u ( e 1 s + e 2 r ) = u ( k Δ + e 1 s + e 2 r ) ug = k(d_0s - b_0r) + u(e_1s + e_2r) = u(k\Delta+ e_1s + e_2r) ug = k ( d 0 s − b 0 r ) + u ( e 1 s + e 2 r ) = u ( k Δ + e 1 s + e 2 r ) . Cancelling u u u in the domain S S S and lifting back, which adds a constant multiple of h h h , proves the assertion.
Proof of Proposition 20.3. By Lemmas 20.4 and 20.5 , only deg A = 6 \deg A = 6 deg A = 6 remains, with height three, and Lemma 20.6 applies. Every entry of G G G vanishes on A A A , so it lies in ( I A ) 2 (I_A)_2 ( I A ) 2 ; as s ∣ U = r ∣ U = 0 s|_U = r|_U = 0 s ∣ U = r ∣ U = 0 , Lemma 20.7 gives constant real symmetric matrices D 1 , D 2 D_1,D_2 D 1 , D 2 with G ( b ) = h ( b ) D 1 + Δ ( b ) D 2 G(b) = h(b)D_1 + \Delta(b)D_2 G ( b ) = h ( b ) D 1 + Δ ( b ) D 2 for b ∈ U b \in U b ∈ U . So the entries of G ∣ U G|_U G ∣ U span at most two forms, and, G ( b ) G(b) G ( b ) being positive semidefinite of rank eight for real b ≠ 0 b \ne0 b = 0 , Lemma 12.2 (with n = dim U = 6 n = \dim U = 6 n = dim U = 6 ) makes Im G ( b ) \operatorname{Im} G(b) Im G ( b ) , hence Im B ( b ) \operatorname{Im} B(b) Im B ( b ) , one fixed eight-space for nonzero b ∈ U b \in U b ∈ U . As dim U = 6 \dim U = 6 dim U = 6 , Lemma 12.5 (1) shows that it is E E E , so K T B ( b ) = 0 K^{T}B(b) = 0 K T B ( b ) = 0 for b ∈ U b \in U b ∈ U . By linearity, K T B ( a ) = u ( a ) Θ 1 + v ( a ) Θ 2 K^{T}B(a)=u(a)\Theta_{1}+v(a)\Theta_{2} K T B ( a ) = u ( a ) Θ 1 + v ( a ) Θ 2 with constant real matrices Θ 1 , Θ 2 \Theta_{1},\Theta_{2} Θ 1 , Θ 2 . Hence the form Q ( a ) = tr ( K T B ( a ) B ( a ) T K ) = tr ( K T G ( a ) K ) Q(a)=\operatorname{tr}(K^{T}B(a)B(a)^{T}K)=\operatorname{tr}(K^{T}G(a)K) Q ( a ) = tr ( K T B ( a ) B ( a ) T K ) = tr ( K T G ( a ) K ) is a binary quadratic form in u , v u,v u , v , and it vanishes on A A A , where G = 0 G=0 G = 0 . If Q ≠ 0 Q\ne0 Q = 0 , it factors over C \mathbb{C} C into two linear forms, and the integral coordinate ring of A A A forces one of them to vanish on A A A , contrary to full complex span. So Q = 0 Q=0 Q = 0 ; on real coefficients it is a sum of squares of real entries, so K T B ( a ) = 0 K^{T}B(a)=0 K T B ( a ) = 0 for every a a a , and rank eight gives (20.1).
A conjugate pair of images
Proposition 20.8. Suppose A A A and A ‾ \overline{A} A are distinct irreducible fourfolds in V ( h , G ) V(h,G) V ( h , G ) whose combined complex span is P 7 \mathbb{P}^{7} P 7 . Then (20.1) holds.
Proof. Let U U U be the vector span of A A A and d = dim C U d=\dim_{\mathbb{C}}U d = dim C U ; as in Lemma 20.1 , 6 ≤ d ≤ 8 6\le d\le8 6 ≤ d ≤ 8 . If d = 8 d=8 d = 8 , both images have degree at least four and Proposition 20.2 applies. If d = 7 d=7 d = 7 , the restriction h ∣ U h|_{U} h ∣ U has rank at least six, and the affine coordinate ring of its quadric is factorial by the Klein–Nagata theorem [29 Theorem 8.2, p. 26] , with a polynomial extension for a possible one-dimensional radical. The cone over A A A is a height-one homogeneous prime, hence generated modulo h ∣ U h|_{U} h ∣ U by one homogeneous element f f f ; full span in U U U excludes deg f = 1 \deg f=1 deg f = 1 , so deg A = 2 deg f ≥ 4 \deg A=2\deg f\ge4 deg A = 2 deg f ≥ 4 and Proposition 20.2 applies again.
Suppose d = 6 d=6 d = 6 . Then h ∣ U h|_{U} h ∣ U has rank at least four, so its null locus is an irreducible quadric fourfold and equals A A A ; so the quadratic entries of G G G restricted to U U U , which vanish on A A A , are constant multiples of h ∣ U h|_{U} h ∣ U : G ( a ) = h ( a ) D G(a)=h(a)D G ( a ) = h ( a ) D for a ∈ U a\in U a ∈ U and one constant complex symmetric matrix D D D . Since U + U ‾ = C 8 U+\overline{U}=\mathbb{C}^{8} U + U = C 8 , the conjugation-invariant space U ∩ U ‾ U\cap\overline{U} U ∩ U has complex dimension four and contains a nonzero real vector w w w . Real rank eight of G ( w ) G(w) G ( w ) forces h ( w ) ≠ 0 h(w)\ne0 h ( w ) = 0 , so D = G ( w ) / h ( w ) D=G(w)/h(w) D = G ( w ) / h ( w ) is real of rank eight; put T = Im R D T=\operatorname{Im}_{\mathbb{R}}D T = Im R D . For complex a ∈ U a\in U a ∈ U with h ( a ) ≠ 0 h(a)\ne0 h ( a ) = 0 the Gram has image T C T_{\mathbb{C}} T C and rank eight, and since B ( a ) B(a) B ( a ) has eight columns, its image is T C T_{\mathbb{C}} T C as well. So the projection of the linear pencil B ∣ U B|_{U} B ∣ U to V C / T C V_{\mathbb{C}}/T_{\mathbb{C}} V C / T C vanishes on a dense open set, hence everywhere, and conjugation gives the same on U ‾ \overline{U} U . As U + U ‾ = C 8 U+\overline{U}=\mathbb{C}^{8} U + U = C 8 , every B ( a ) B(a) B ( a ) maps into this fixed eight-space, and the radial columns identify its real form with E E E .
Remark 20.9. For a residue image of type ( α , s ) = ( 4 , 4 ) (\alpha,s)=(4,4) ( α , s ) = ( 4 , 4 ) , the rank identity rank G = 2 s − 8 \operatorname{rank}G=2s-8 rank G = 2 s − 8 gives the vanishing of G G G on the image, as required above; full real support becomes full complex span when the image is geometrically integral and real-defined. For a simple real-irreducible quadratic or quartic pole that splits over C \mathbb{C} C with α = s = 4 \alpha=s=4 α = s = 4 , Lemma 17.6 makes the two geometric components give distinct images, and full real residue support gives their combined span.
The divisor budget for five-dimensional residue images
Throughout this section ν = 8 \nu=8 ν = 8 , the anchor is regular, B ( a ) B(a) B ( a ) is the transpose pencil of Setting (T), with coefficients a = ( a 1 , … , a 8 ) a=(a_{1},\ldots,a_{8}) a = ( a 1 , … , a 8 ) , and h ( a ) = a T H a h(a)=a^{T}Ha h ( a ) = a T H a is the nondegenerate indefinite coefficient form. By the isotropy in (10.9), every residue image lies in the quadric V ( h ) ⊂ P C 7 V(h)\subset\mathbb{P}^{7}_{\mathbb{C}} V ( h ) ⊂ P C 7 . A pole has type ( α , s ) = ( 5 , 5 ) (\alpha,s)=(5,5) ( α , s ) = ( 5 , 5 ) when each geometric pole component has an irreducible closed residue image A A A of dimension five on which B B B has generic rank five; distinct geometric pole components have distinct images (Lemma 17.6 ). Since rank is lower semicontinuous, rank B ( a ) ≤ 5 \operatorname{rank}B(a)\le5 rank B ( a ) ≤ 5 at every point of A A A , so A A A has no real point.
The real residue support (as in § 17 ) is the smallest real subspace U ⊂ R 8 U\subset\mathbb{R}^{8} U ⊂ R 8 whose complexification contains every residue image. A real-defined image in a complex hyperplane { c ⋅ a = 0 } \{c\cdot a=0\} { c ⋅ a = 0 } also lies in { c ‾ ⋅ a = 0 } \{\overline{c}\cdot a=0\} { c ⋅ a = 0 } , hence in a real hyperplane; so a real-defined image of a component with real support 8 8 8 spans P 7 \mathbb{P}^{7} P 7 .
The budget
Let S = C [ a 1 , … , a 8 ] / ( h ) S = \mathbb{C}[a_1,\ldots,a_8]/(h) S = C [ a 1 , … , a 8 ] / ( h ) with its standard grading. By the Klein–Nagata theorem [29 Theorem 8.2, p. 26] , a nondegenerate quadratic form in at least five variables over a field of characteristic not two gives a factorial quotient, so S S S is a UFD. The affine cone over a five-dimensional irreducible A ⊂ V ( h ) A \subset V(h) A ⊂ V ( h ) is a height-one prime of S S S , so its ideal is generated by a homogeneous irreducible f A f_A f A of some degree e A ≥ 1 e_A \ge1 e A ≥ 1 ; if A A A spans P 7 \mathbb{P}^7 P 7 , then e A ≥ 2 e_A \ge2 e A ≥ 2 .
Lemma 21.1 (Divisor budget). Let A 1 , … , A k A_1,\ldots,A_k A 1 , … , A k be distinct residue images of type ( 5 , 5 ) (5,5) ( 5 , 5 ) . Then e A 1 + ⋯ + e A k ≤ 2 e_{A_1}+\cdots+e_{A_k}\le2 e A 1 + ⋯ + e A k ≤ 2 .
Proof. Over the fraction field of S S S the matrix B B B has rank 8 8 8 : the real cone of h h h contains nonzero points, where B B B has rank 8 8 8 , so some maximal minor of B B B is nonzero in S S S . For each i i i , localizing S S S at the height-one prime ( f A i ) (f_{A_i}) ( f A i ) gives a discrete valuation ring. The reduction of B B B has rank five at the generic point of A i A_i A i , so in the local Smith form of the 15 × 8 15\times8 15 × 8 matrix B B B (its image lies in X ⊥ X^\perp X ⊥ ) at least three of the eight invariant factors lie in the maximal ideal, and every maximal minor of B B B is divisible in S S S by f A i 3 f_{A_i}^3 f A i 3 . The f A i f_{A_i} f A i are pairwise nonassociate primes, so every maximal minor is divisible by ∏ i f A i 3 \prod_i f_{A_i}^3 ∏ i f A i 3 , of degree 3 ∑ i e A i 3\sum_i e_{A_i} 3 ∑ i e A i . A maximal minor has degree 8 8 8 and some maximal minor is nonzero in S S S , so 3 ∑ i e A i ≤ 8 3\sum_i e_{A_i}\le8 3 ∑ i e A i ≤ 8 . □
Consequences
Proposition 21.2 (Two quadratic poles). Suppose that the reduced denominator is Q 1 Q 2 Q_1Q_2 Q 1 Q 2 with two distinct real quadratic factors, that every real residue support is 8 8 8 , and that all geometric pole components have type ( 5 , 5 ) (5,5) ( 5 , 5 ) . This configuration does not occur.
Proof. Each real quadratic factor contributes at least two to the sum of the degrees e e e . If it is geometrically irreducible, its single image is real-defined of full span, so e ≥ 2 e\ge2 e ≥ 2 ; if it splits, its two conjugate geometric components have distinct images, each with e ≥ 1 e\ge1 e ≥ 1 . Images of different pole components are distinct by Lemma 17.6 . So the two factors, of either kind, contribute at least four, contrary to Lemma 21.1 . □
Proposition 21.3 (A single real pole). Suppose that the pole has one real component, with full real support and type ( 5 , 5 ) (5,5) ( 5 , 5 ) .
If its image A A A is real-defined and geometrically integral, then A = V ( h , f ) A=V(h,f) A = V ( h , f ) for a real quadric f f f ; it is a complete intersection of degree four with no real point.
Otherwise there are exactly two images, A A A and A ‾ \overline{A} A , and A = V ( h , ℓ ) A=V(h,\ell) A = V ( h , ℓ ) , A ‾ = V ( h , ℓ ‾ ) \overline{A}=V(h,\overline{\ell}) A = V ( h , ℓ ) for a complex linear form ℓ \ell ℓ that is not proportional to a real form.
Proof. A real irreducible polynomial is geometrically irreducible or splits into one pair of complex conjugate factors.
(1) Full real support gives full complex span (as shown above), so e A ≥ 2 e_A\ge2 e A ≥ 2 , and Lemma 21.1 gives e A = 2 e_A=2 e A = 2 : A = V ( h , f A ) A=V(h,f_A) A = V ( h , f A ) with f A f_A f A of degree two. Since A A A is real-defined, f A ‾ \overline{f_A} f A generates the same prime, so f A ‾ = c f A \overline{f_A}=cf_A f A = c f A with c c c a nonzero constant (the only units of S S S in degree zero). Conjugating twice gives ∣ c ∣ = 1 |c|=1 ∣ c ∣ = 1 ; writing c = ω / ω ‾ c=\omega/\overline{\omega} c = ω / ω , the class ω f A \omega f_A ω f A is fixed by conjugation, and any real lift f f f of it defines the same complete intersection, of degree 2 ⋅ 2 = 4 2\cdot2=4 2 ⋅ 2 = 4 ; like every image, it has no real point.
(2) The two images are distinct by Lemma 17.6 and conjugate, so e A = e A ‾ e_A=e_{\overline{A}} e A = e A , and Lemma 21.1 gives e A = e A ‾ = 1 e_A=e_{\overline{A}}=1 e A = e A = 1 , so A = V ( h , ℓ ) A=V(h,\ell) A = V ( h , ℓ ) with ℓ \ell ℓ linear. If ℓ \ell ℓ were proportional to a real form, A A A would be real-defined and equal to A ‾ \overline{A} A . □
The conjugate pair: definiteness and the positive model
In case (2) of Proposition 21.3 write ℓ = ℓ 1 + i ℓ 2 \ell=\ell_1+i\ell_2 ℓ = ℓ 1 + i ℓ 2 with real linear forms ℓ 1 , ℓ 2 \ell_1,\ell_2 ℓ 1 , ℓ 2 , linearly independent because ℓ \ell ℓ is not proportional to a real form, and put
W = { a ∈ R 8 : ℓ 1 ( a ) = ℓ 2 ( a ) = 0 } , r = ℓ 1 2 + ℓ 2 2 = ℓ ℓ ˉ . W=\{a\in\mathbb{R}^{8}:\ell_1(a)=\ell_2(a)=0\},\qquad r=\ell_1^2+\ell_2^2=\ell\bar{\ell}. W = { a ∈ R 8 : ℓ 1 ( a ) = ℓ 2 ( a ) = 0 } , r = ℓ 1 2 + ℓ 2 2 = ℓ ℓ ˉ .
Then dim W = 6 \dim W=6 dim W = 6 , and r r r is a real positive semidefinite form of rank two.
Lemma 21.4. The restriction h ∣ W h|_W h ∣ W is definite. Consequently the signature of h h h is ( 6 , 2 ) (6,2) ( 6 , 2 ) or ( 7 , 1 ) (7,1) ( 7 , 1 ) , up to an overall sign.
Proof. A real point of A = V ( h , ℓ ) A=V(h,\ell) A = V ( h , ℓ ) is a real a ≠ 0 a\ne0 a = 0 with h ( a ) = 0 h(a)=0 h ( a ) = 0 and ℓ 1 ( a ) = ℓ 2 ( a ) = 0 \ell_1(a)=\ell_2(a)=0 ℓ 1 ( a ) = ℓ 2 ( a ) = 0 , that is, a real zero of h h h on W W W . As A A A has no real point, h ∣ W h|_W h ∣ W has no nonzero real zero, so it is definite. A nondegenerate form that is definite on a six-dimensional subspace has at least six squares of one sign; since h h h is indefinite, its signature is ( 6 , 2 ) (6,2) ( 6 , 2 ) or ( 7 , 1 ) (7,1) ( 7 , 1 ) up to sign.
Proposition 21.5 (Positive model of the conjugate pair). There are a positive definite real quadratic form q q q and a real constant λ ≠ 0 \lambda\ne0 λ = 0 such that ρ = h − λ q \rho=h-\lambda q ρ = h − λ q is a semidefinite form of rank two with radical W W W , whose restriction to W ⊥ q W^{\perp_q} W ⊥ q is definite. In q q q -orthonormal coordinates a = ( x , z ) a=(x,z) a = ( x , z ) with x ∈ W x\in W x ∈ W and z ∈ W ⊥ q z\in W^{\perp_q} z ∈ W ⊥ q ,
q = ∣ x ∣ 2 + ∣ z ∣ 2 , h = λ q + ρ ( z ) , A ∪ A ‾ = V ( h , r ) = V ( h , q ) = V ( q , ρ ) . q=|x|^2+|z|^2,\qquad h=\lambda q+\rho(z),\qquad A\cup\overline{A}=V(h,r)=V(h,q)=V(q,\rho). q = ∣ x ∣ 2 + ∣ z ∣ 2 , h = λ q + ρ ( z ) , A ∪ A = V ( h , r ) = V ( h , q ) = V ( q , ρ ) . In particular the pencil spanned by h h h and q q q contains the rank-two member ρ \rho ρ , and λ \lambda λ is a relative eigenvalue of h h h with respect to q q q of multiplicity six.
Proof. The forms h h h and r r r have no common nonzero real zero: such a zero would lie in W W W and be a real zero of h h h there, which Lemma 21.4 excludes. As they are forms in 8 ≥ 3 8\ge3 8 ≥ 3 variables, the theorem of Finsler and Calabi [7 , 12 ] gives a positive definite combination q = c 1 h + c 2 r q=c_1h+c_2r q = c 1 h + c 2 r . Here c 1 ≠ 0 c_1\ne0 c 1 = 0 because r r r has rank two, and c 2 ≠ 0 c_2\ne0 c 2 = 0 because h h h is indefinite. Put λ = 1 / c 1 \lambda=1/c_1 λ = 1/ c 1 and ρ = − ( c 2 / c 1 ) r = h − λ q \rho=-(c_2/c_1)r=h-\lambda q ρ = − ( c 2 / c 1 ) r = h − λ q , a nonzero multiple of r r r ; so ρ \rho ρ is semidefinite of rank two with radical W W W , and definite on W ⊥ q W^{\perp_q} W ⊥ q .
In q q q -orthonormal coordinates x x x on W W W and z z z on W ⊥ q W^{\perp_q} W ⊥ q , the forms ℓ 1 , ℓ 2 \ell_1,\ell_2 ℓ 1 , ℓ 2 vanish on W W W , so they, and hence ρ \rho ρ , depend on z z z alone; this gives the model. Since c 2 ≠ 0 c_2\ne0 c 2 = 0 , V ( h , q ) = V ( h , r ) = V ( h , ℓ ℓ ˉ ) = V ( h , ℓ ) ∪ V ( h , ℓ ˉ ) V(h,q)=V(h,r)=V(h,\ell\bar{\ell})=V(h,\ell)\cup V(h,\bar{\ell}) V ( h , q ) = V ( h , r ) = V ( h , ℓ ℓ ˉ ) = V ( h , ℓ ) ∪ V ( h , ℓ ˉ ) , and V ( h , q ) = V ( q , h − λ q ) = V ( q , ρ ) V(h,q)=V(q,h-\lambda q)=V(q,\rho) V ( h , q ) = V ( q , h − λ q ) = V ( q , ρ ) . On W W W the relative eigenvalue of h h h with respect to q q q is λ \lambda λ , and on W ⊥ q W^{\perp_q} W ⊥ q the form h − λ q = ρ h-\lambda q=\rho h − λ q = ρ is nondegenerate, so λ \lambda λ has multiplicity exactly six.
Remark 21.6. Each component V ( h , ℓ ) V(h,\ell) V ( h , ℓ ) is a quadric of rank six or seven in its hyperplane, since restricting the rank-eight form h h h to a hyperplane lowers its rank by at most two; in particular it is irreducible and spans its hyperplane. Proposition 21.5 gives the normal form (25.1) used in Sections 25 and 26, where the conjugate pair, a reducible complete intersection, is treated.
Alignment criteria for quadratic normal forms
We work in Setting (T) of § 12 , with all eight coefficient directions; over C \mathbb{C} C , the transpose is that of the complex-bilinear extension. As in § 10 , write B ( a ) g = N g a B(a)g=N_ga B ( a ) g = N g a . Since rank B ( a ) = 8 \operatorname{rank}B(a)=8 rank B ( a ) = 8 for a ≠ 0 a\ne0 a = 0 ,
N g : R 8 ⟶ V is injective for every g ≠ 0. (54) N_g:\mathbb{R}^{8}\longrightarrow V\text{ is injective for every }g\ne0. \tag*{(54)} N g : R 8 ⟶ V is injective for every g = 0. (54)
Put x 0 = K g 0 x_0=Kg_0 x 0 = K g 0 . As K T B ( a ) g 0 = − K T L a = 0 K^{T}B(a)g_0=-K^{T}La=0 K T B ( a ) g 0 = − K T L a = 0 , skewness makes g 0 g_0 g 0 orthogonal to the image of K T B ( a ) K^{T}B(a) K T B ( a ) , so
Im B ( a ) ⊂ x 0 ⊥ , ⟨ K g , B ( a ) g ⟩ = 0. (55) \operatorname{Im}B(a)\subset x_0^{\perp},\qquad\langle Kg,B(a)g\rangle=0. \tag*{(55)} Im B ( a ) ⊂ x 0 ⊥ , ⟨ K g , B ( a ) g ⟩ = 0. (55)
Balanced normal forms
Proposition 22.1. Suppose that q q q is a positive definite quadratic form on R 8 \mathbb{R}^{8} R 8 and h h h is a real quadratic form, with eigenvalues λ 1 ≤ ⋯ ≤ λ 8 \lambda_{1} \le\cdots\le\lambda_{8} λ 1 ≤ ⋯ ≤ λ 8 relative to q q q . Assume there are constant symmetric endomorphisms D 0 , D 1 D_{0},D_{1} D 0 , D 1 of V V V and real linear maps u , v : R 8 → V u,v:\mathbb{R}^{8} \to V u , v : R 8 → V such that
G ( a ) = h ( a ) D 0 + q ( a ) D 1 + u ( a ) u ( a ) T − v ( a ) v ( a ) T (56) G(a) = h(a)D_{0} + q(a)D_{1} + u(a)u(a)^{T} - v(a)v(a)^{T} \tag*{(56)} G ( a ) = h ( a ) D 0 + q ( a ) D 1 + u ( a ) u ( a ) T − v ( a ) v ( a ) T (56) (a balanced normal form). If λ 2 < λ 7 \lambda_{2} < \lambda_{7} λ 2 < λ 7 , then Im B ( a ) = E \operatorname{Im} B(a) = E Im B ( a ) = E for every nonzero real a a a .
Proof. Write D ( t ) = D 1 + t D 0 D(t) = D_{1} + tD_{0} D ( t ) = D 1 + t D 0 . Fix b ∈ V b \in V b ∈ V ; the solution space A b ⊂ R 8 A_{b} \subset\mathbb{R}^{8} A b ⊂ R 8 of ⟨ b , u ( a ) ⟩ = 0 \langle b,u(a)\rangle= 0 ⟨ b , u ( a )⟩ = 0 has codimension at most one, and on A b ∖ { 0 } A_{b} \setminus\{0\} A b ∖ { 0 } positivity of (56) gives q ( a ) ⟨ D ( h ( a ) / q ( a ) ) b , b ⟩ ≥ ⟨ b , v ( a ) ⟩ 2 ≥ 0 q(a)\langle D(h(a)/q(a))b,b\rangle\ge\langle b,v(a)\rangle^{2} \ge0 q ( a ) ⟨ D ( h ( a ) / q ( a )) b , b ⟩ ≥ ⟨ b , v ( a ) ⟩ 2 ≥ 0 . By the min–max principle the minimum of h / q h/q h / q on A b A_{b} A b is at most λ 2 \lambda_{2} λ 2 and its maximum is at least λ 7 \lambda_{7} λ 7 , and its range is an interval; as b b b was arbitrary, D ( t ) ⪰ 0 D(t) \succeq0 D ( t ) ⪰ 0 for every t ∈ [ λ 2 , λ 7 ] t \in[\lambda_{2},\lambda_{7}] t ∈ [ λ 2 , λ 7 ] (the test vector a a a may depend on b b b ). By Lemma 12.1 , D ( t ) D(t) D ( t ) has one image F \mathcal{F} F for all interior t t t ; let r = dim F r = \dim\mathcal{F} r = dim F . The set S = { a ≠ 0 : λ 2 < h ( a ) / q ( a ) < λ 7 } S = \{a \ne0 : \lambda_{2} < h(a)/q(a) < \lambda_{7}\} S = { a = 0 : λ 2 < h ( a ) / q ( a ) < λ 7 } is nonempty and open. For a ∈ S a \in S a ∈ S put G + ( a ) = q ( a ) D ( h ( a ) / q ( a ) ) + u ( a ) u ( a ) T G_{+}(a) = q(a)D(h(a)/q(a)) + u(a)u(a)^{T} G + ( a ) = q ( a ) D ( h ( a ) / q ( a )) + u ( a ) u ( a ) T . Both G + ( a ) G_{+}(a) G + ( a ) and G ( a ) = G + ( a ) − v ( a ) v ( a ) T G(a) = G_{+}(a) - v(a)v(a)^{T} G ( a ) = G + ( a ) − v ( a ) v ( a ) T are positive semidefinite; testing a vector in ker G + ( a ) \operatorname{ker}G_{+}(a) ker G + ( a ) shows v ( a ) ∈ Im G + ( a ) v(a) \in\operatorname{Im}G_{+}(a) v ( a ) ∈ Im G + ( a ) . As G ( a ) G(a) G ( a ) has rank eight, rank G + ( a ) \operatorname{rank}G_{+}(a) rank G + ( a ) is eight or nine, and r ≤ rank G + ( a ) ≤ r + 1 r \le\operatorname{rank}G_{+}(a) \le r + 1 r ≤ rank G + ( a ) ≤ r + 1 , so 7 ≤ r ≤ 9 7 \le r \le9 7 ≤ r ≤ 9 .
If r = 9 r = 9 r = 9 , then u ( a ) ∈ F u(a) \in\mathcal{F} u ( a ) ∈ F for a ∈ S a \in S a ∈ S , since otherwise G + ( a ) G_{+}(a) G + ( a ) would have rank ten and subtracting a rank-one matrix could not give rank eight. So v ( a ) ∈ F v(a) \in\mathcal{F} v ( a ) ∈ F and Im G ( a ) ⊂ F \operatorname{Im}G(a) \subset\mathcal{F} Im G ( a ) ⊂ F ; over the reals Im G ( a ) = Im B ( a ) \operatorname{Im}G(a) = \operatorname{Im}B(a) Im G ( a ) = Im B ( a ) , and since B B B is linear its image lies in F \mathcal{F} F for every a a a . Lemma 12.5 (2) with U = R 8 U = \mathbb{R}^{8} U = R 8 applies.
Suppose r ∈ { 7 , 8 } r \in\{7,8\} r ∈ { 7 , 8 } , and let π : V → F ⊥ \pi: V \to\mathcal{F}^{\perp} π : V → F ⊥ be the orthogonal projection. For a ∈ S a \in S a ∈ S ,
( π B ( a ) ) ( π B ( a ) ) T = ( π u ( a ) ) ( π u ( a ) ) T − ( π v ( a ) ) ( π v ( a ) ) T ⪰ 0. (\pi B(a))(\pi B(a))^{T} = (\pi u(a))(\pi u(a))^{T} - (\pi v(a))(\pi v(a))^{T} \succeq0. ( π B ( a )) ( π B ( a ) ) T = ( π u ( a )) ( π u ( a ) ) T − ( π v ( a )) ( π v ( a ) ) T ⪰ 0.
A positive semidefinite difference of two matrices of rank at most one has rank at most one, since a vector orthogonal to π u ( a ) \pi u(a) π u ( a ) is then orthogonal to π v ( a ) \pi v(a) π v ( a ) . So every two-minor of the linear matrix π B ( a ) \pi B(a) π B ( a ) vanishes on S S S , hence identically, and by Lemma 12.6 the matrices π B ( a ) \pi B(a) π B ( a ) have a common image line or a common row line.
In the common-image case the images of B B B lie in F \mathcal{F} F plus one fixed line, a space of dimension at most nine, and Lemma 12.5 (2) applies again. In the common-row case there is a fixed nonzero covector e ∗ e^{*} e ∗ on the g g g -space R 8 \mathbb{R}^{8} R 8 (the domain of B ( a ) B(a) B ( a ) ) with π B ( a ) = w ( a ) e ∗ \pi B(a) = w(a)e^{*} π B ( a ) = w ( a ) e ∗ , so N g ( R 8 ) ⊂ F N_{g}(\mathbb{R}^{8}) \subset\mathcal{F} N g ( R 8 ) ⊂ F for every nonzero g ∈ ker e ∗ g \in\operatorname{ker}e^{*} g ∈ ker e ∗ . Injectivity (54) excludes r = 7 r = 7 r = 7 and gives N g ( R 8 ) = F N_{g}(\mathbb{R}^{8}) = \mathcal{F} N g ( R 8 ) = F when r = 8 r = 8 r = 8 , and (55) gives F ⊥ ⊥ K ( ker e ∗ ) \mathcal{F}^{\perp} \perp K(\operatorname{ker}e^{*}) F ⊥ ⊥ K ( ker e ∗ ) and F ⊥ ⊥ x 0 \mathcal{F}^{\perp} \perp x_{0} F ⊥ ⊥ x 0 . If g 0 ∉ ker e ∗ g_{0} \notin\operatorname{ker}e^{*} g 0 ∈ / ker e ∗ , these spaces span E ⊥ E^{\perp} E ⊥ , so F = E \mathcal{F} = E F = E ; if g 0 ∈ ker e ∗ g_{0} \in\operatorname{ker}e^{*} g 0 ∈ ker e ∗ , the identity N g 0 = − L N_{g_{0}} = -L N g 0 = − L gives F = E \mathcal{F} = E F = E directly. In either case π = K K T \pi= KK^{T} π = K K T , so the skew map K T B ( a ) K^{T}B(a) K T B ( a ) has rank at most one and is therefore zero, and rank eight gives the result. ∎
Proposition 22.2. Let q , h q,h q , h be as in Proposition 22.1 and suppose that (56) holds. Then Im B ( a ) = E \operatorname{Im} B(a) = E Im B ( a ) = E for every nonzero real a a a in either of the following situations:
Every eigenvalue of h h h relative to q q q has multiplicity at most five.
There is a q q q -orthogonal decomposition R 8 = W ⊕ Z \mathbb{R}^{8} = W \oplus Z R 8 = W ⊕ Z , with dim W = 6 \dim W = 6 dim W = 6 , dim Z = 2 \dim Z = 2 dim Z = 2 , such that h = λ q + ρ h = \lambda q + \rho h = λ q + ρ , where ρ \rho ρ vanishes on W W W and its restriction to Z Z Z is definite.
Proof. Equality λ 2 = λ 7 \lambda_{2} = \lambda_{7} λ 2 = λ 7 would require an eigenvalue of multiplicity at least six, which the first situation excludes. In the second, λ \lambda λ is an extreme eigenvalue of multiplicity six: if ρ ∣ Z \rho|_{Z} ρ ∣ Z is positive definite, λ 2 = λ < λ 7 \lambda_{2} = \lambda< \lambda_{7} λ 2 = λ < λ 7 , and if it is negative definite, λ 2 < λ = λ 7 \lambda_{2} < \lambda= \lambda_{7} λ 2 < λ = λ 7 . In both situations λ 2 < λ 7 \lambda_{2} < \lambda_{7} λ 2 < λ 7 , and Proposition 22.1 applies. ∎
:::
Unbalanced forms and vanishing on a stratum
Proposition 22.3 (Unbalanced normal form). Let h , q h,q h , q be independent real quadratic forms on R 8 \mathbb{R}^{8} R 8 , with q q q positive definite. Suppose that D 0 , D 1 D_0,D_1 D 0 , D 1 are constant symmetric endomorphisms of V V V , u ∈ V u \in V u ∈ V is a fixed nonzero vector, and b : R 8 → V b:\mathbb{R}^{8}\to V b : R 8 → V is quadratic, such that
G ( a ) = h ( a ) D 0 + q ( a ) D 1 + u b ( a ) T + b ( a ) u T G(a)=h(a)D_0+q(a)D_1+u b(a)^{T}+b(a)u^{T} G ( a ) = h ( a ) D 0 + q ( a ) D 1 + u b ( a ) T + b ( a ) u T (an unbalanced normal form). Then Im B ( a ) = E \operatorname{Im} B(a)=E Im B ( a ) = E for every nonzero real a a a .
Proof. Compress to u ⊥ u^\perp u ⊥ : the term u b T + b u T u b^T+b u^T u b T + b u T compresses to zero, so the compression is G ⊥ ( a ) = h ( a ) D 0 ⊥ + q ( a ) D 1 ⊥ G_\perp(a)=h(a)D^\perp_0+q(a)D^\perp_1 G ⊥ ( a ) = h ( a ) D 0 ⊥ + q ( a ) D 1 ⊥ , where D i ⊥ D^\perp_i D i ⊥ are the compressions of D i D_i D i ; it is positive semidefinite of rank at most eight. On q = 1 q=1 q = 1 the quotient h / q h/q h / q ranges over a nondegenerate closed interval I I I , so D 1 ⊥ + t D 0 ⊥ ⪰ 0 D^\perp_1+tD^\perp_0\succeq0 D 1 ⊥ + t D 0 ⊥ ⪰ 0 for t ∈ I t\in I t ∈ I , and by Lemma 12.1 its image for interior t t t is F = Im D 0 ⊥ + Im D 1 ⊥ F=\operatorname{Im}D^\perp_0+\operatorname{Im}D^\perp_1 F = Im D 0 ⊥ + Im D 1 ⊥ , of dimension at most eight. Hence Im G ⊥ ( a ) ⊂ F \operatorname{Im}G_\perp(a)\subset F Im G ⊥ ( a ) ⊂ F for every a a a , endpoint values included. Positivity of G ( a ) G(a) G ( a ) puts its cross block in Im G ⊥ ( a ) \operatorname{Im}G_\perp(a) Im G ⊥ ( a ) , since a vector in the kernel of a positive semidefinite principal block is annihilated by the corresponding cross block; so Im G ( a ) ⊂ R u + F \operatorname{Im}G(a)\subset\mathbb{R}u+F Im G ( a ) ⊂ R u + F . Over the reals Im G ( a ) = Im B ( a ) \operatorname{Im}G(a)=\operatorname{Im}B(a) Im G ( a ) = Im B ( a ) , and dim ( R u + F ) ≤ 9 \dim(\mathbb{R}u+F)\le9 dim ( R u + F ) ≤ 9 , so Lemma 12.5 (2) with U = R 8 U=\mathbb{R}^{8} U = R 8 applies. □
For the stratum statement, choose coefficient coordinates R 8 = W ⊕ Z \mathbb{R}^{8}=W\oplus Z R 8 = W ⊕ Z with dim W = m ∈ { 4 , 5 } \dim W=m\in\{4,5\} dim W = m ∈ { 4 , 5 } and
q ( x , z ) = ∣ x ∣ 2 + ∣ z ∣ 2 , h ( x , z ) = λ q ( x , z ) + ρ ( z ) , q(x,z)=|x|^2+|z|^2,\qquad h(x,z)=\lambda q(x,z)+\rho(z), q ( x , z ) = ∣ x ∣ 2 + ∣ z ∣ 2 , h ( x , z ) = λ q ( x , z ) + ρ ( z ) ,
where ρ \rho ρ is a nondegenerate real quadratic form on Z Z Z of any signature. Let
A = { ( x , z ) ∈ C 8 : q ( x , z ) = ρ ( z ) = 0 } , Σ = { ( x , 0 ) : ∣ x ∣ 2 = 0 } , \mathcal{A}=\{(x,z)\in\mathbb{C}^{8}:q(x,z)=\rho(z)=0\},\qquad\Sigma=\{(x,0):|x|^2=0\}, A = {( x , z ) ∈ C 8 : q ( x , z ) = ρ ( z ) = 0 } , Σ = {( x , 0 ) : ∣ x ∣ 2 = 0 } ,
and assume rank G ≤ 2 \operatorname{rank}G\le2 rank G ≤ 2 at every point of A \mathcal{A} A , singular points included. In the residue construction this follows from the generic rank bound on every component, since the three-minors are polynomials.
Proposition 22.4 (Vanishing on a multiplicity-four or -five stratum). In this setting, if G G G vanishes on Σ \Sigma Σ , then Im B ( a ) = E \operatorname{Im}B(a)=E Im B ( a ) = E for every nonzero real a a a .
Proof. Lemma 12.9 applies with w = m w=m w = m : its hypotheses (i) and (ii) are the vanishing on Σ \Sigma Σ and the rank bound on A = V ( q , ρ ) \mathcal{A}=V(q,\rho) A = V ( q , ρ ) . If m = 5 m=5 m = 5 , then dim Z = 3 \dim Z=3 dim Z = 3 and Lemma 12.9 (3) gives the result.
Let m = 4 m=4 m = 4 , let T T T be the skew linear map of Lemma 12.9 (1), with K T B ( x , z ) = T ( z ) K^{T}B(x,z)=T(z) K T B ( x , z ) = T ( z ) , and let Γ 1 , Γ 2 , Γ 3 \Gamma_1,\Gamma_2,\Gamma_3 Γ 1 , Γ 2 , Γ 3 be its operators for an orthonormal basis of the orthogonal complement of x ∗ x_* x ∗ in W W W . For a complex null vector z 0 z_0 z 0 of ρ \rho ρ with T ( z 0 ) ≠ 0 T(z_0)\ne0 T ( z 0 ) = 0 , Lemma 12.9 (2) gives a two-dimensional image U = Im T ( z 0 ) T U=\operatorname{Im}T(z_0)^T U = Im T ( z 0 ) T , invariant under each Γ i \Gamma_i Γ i , with tr ( Γ i ∣ U ) = 0 \operatorname{tr}(\Gamma_i|_U)=0 tr ( Γ i ∣ U ) = 0 . A skew map with image U U U corresponds to a bivector in Λ 2 U \Lambda^2 U Λ 2 U , on which Γ i \Gamma_i Γ i acts by its trace, that is, by zero; in matrix terms
Γ i T ( z 0 ) + T ( z 0 ) Γ i T = Γ i T ( z 0 ) − T ( z 0 ) Γ i = 0. \Gamma_iT(z_0)+T(z_0)\Gamma_i^T=\Gamma_iT(z_0)-T(z_0)\Gamma_i=0. Γ i T ( z 0 ) + T ( z 0 ) Γ i T = Γ i T ( z 0 ) − T ( z 0 ) Γ i = 0.
This also holds when T ( z 0 ) = 0 T(z_0)=0 T ( z 0 ) = 0 . Each commutator is linear in z z z , and the null cone of the nondegenerate rank-four form ρ \rho ρ spans Z C Z_{\mathbb{C}} Z C , so T ( z ) T(z) T ( z ) commutes with all three Γ i \Gamma_i Γ i for every z z z . The radial identity in (32) gives T ( z ) g 0 = 0 T(z)g_0=0 T ( z ) g 0 = 0 , and commutation with Γ 1 , Γ 2 \Gamma_1,\Gamma_2 Γ 1 , Γ 2 makes T ( z ) T(z) T ( z ) vanish on the real four-space X 0 = span R { g 0 , Γ 1 g 0 , Γ 2 g 0 , Γ 1 Γ 2 g 0 } X_0=\operatorname{span}_{\mathbb{R}}\{g_0,\Gamma_1g_0,\Gamma_2g_0,\Gamma_1\Gamma_2g_0\} X 0 = span R { g 0 , Γ 1 g 0 , Γ 2 g 0 , Γ 1 Γ 2 g 0 } , whose spanning vectors divided by ∥ g 0 ∥ \|g_0\| ∥ g 0 ∥ are orthonormal by the relations of Lemma 12.9 (1). The spaces X 0 X_0 X 0 and X 0 ⊥ X_0^\perp X 0 ⊥ are invariant under Γ 1 , Γ 2 \Gamma_1,\Gamma_2 Γ 1 , Γ 2 , and skewness makes T T T preserve X 0 ⊥ X_0^\perp X 0 ⊥ . On this four-dimensional complement Γ 1 , Γ 2 \Gamma_1,\Gamma_2 Γ 1 , Γ 2 give the left action of the quaternions on a quaternionic line; a commuting real endomorphism is right multiplication by a quaternion, and skew-adjointness makes the quaternion imaginary. Explicitly, for a unit v ∈ X 0 ⊥ v \in\mathcal{X}_0^\perp v ∈ X 0 ⊥ a commuting map is determined by its value at v v v in the basis v , Γ 1 v , Γ 2 v , Γ 1 Γ 2 v v,\Gamma_1v,\Gamma_2v,\Gamma_1\Gamma_2v v , Γ 1 v , Γ 2 v , Γ 1 Γ 2 v , and skewness removes its v v v coefficient. Hence there are real linear forms t 1 , t 2 , t 3 t_1,t_2,t_3 t 1 , t 2 , t 3 on Z Z Z with
T ( z ) ∣ X 0 ⊥ = − τ ( z ) J , τ = t 1 2 + t 2 2 + t 3 2 . (57) \left.T(z)\right|_{\mathcal{X}_0^\perp}=-\tau(z)J,\qquad\tau=t_1^2+t_2^2+t_3^2. \tag*{(57)} T ( z ) ∣ X 0 ⊥ = − τ ( z ) J , τ = t 1 2 + t 2 2 + t 3 2 . (57)
By Lemma 12.9(2), rank T ( z ) ≤ 2 T(z) \le2 T ( z ) ≤ 2 whenever ρ ( z ) = 0 \rho(z)=0 ρ ( z ) = 0 over C \mathbb{C} C . If τ ( z ) ≠ 0 \tau(z) \ne0 τ ( z ) = 0 at such a point, (57) would make T ( z ) T(z) T ( z ) invertible on X 0 ⊥ \mathcal{X}_0^\perp X 0 ⊥ , of rank four. So τ \tau τ vanishes on the irreducible rank-four quadric ρ = 0 \rho=0 ρ = 0 ; both have degree two, so τ = c ρ \tau=c\rho τ = c ρ with c c c real, and rank τ ≤ 3 < 4 = rank ρ \tau\le3 < 4 = \operatorname{rank}\rho τ ≤ 3 < 4 = rank ρ forces c = 0 c=0 c = 0 . A real sum of three squares vanishes identically only if each t i t_i t i is zero, so T = 0 T=0 T = 0 . The lower projection of B B B is therefore zero, and rank eight gives Im B ( a ) = E \operatorname{Im} B(a)=E Im B ( a ) = E .
Positive quadratic models and the factorial case
Throughout, ν = 8 \nu=8 ν = 8 and the anchor is regular (§ 10 ). By Proposition 21.3(1), a real-defined, geometrically integral, full-span residue fivefold is a complete intersection A = V ( h , f ) ⊂ P R 7 A=V(h,f)\subset\mathbb{P}_{\mathbb{R}}^7 A = V ( h , f ) ⊂ P R 7 of two real quadrics, with h h h nondegenerate and indefinite, and A A A has no real point.
Proposition 23.1. There is a positive definite real quadric q q q with span R { h , f } = span R { h , q } \operatorname{span}_{\mathbb{R}}\{h,f\}=\operatorname{span}_{\mathbb{R}}\{h,q\} span R { h , f } = span R { h , q } . Every eigenvalue of h h h relative to q q q has multiplicity at most five. Both A A A and its affine cone are normal.
Proof. The forms h h h and f f f have no common nonzero real zero, since A A A has no real point. As they are forms in $8 \ge 3$ variables, the theorem of Finsler and Calabi [7 , 12 ] gives a positive definite combination q = c 1 h + c 2 f q=c_1h+c_2f q = c 1 h + c 2 f . Indefiniteness of h h h forces $c_2 \ne 0.
In real coordinates with q = ∑ i = 1 8 a i 2 q=\sum_{i=1}^{8}a_i^2 q = ∑ i = 1 8 a i 2 and h = ∑ i = 1 8 λ i a i 2 h=\sum_{i=1}^{8}\lambda_i a_i^2 h = ∑ i = 1 8 λ i a i 2 , the λ i \lambda_i λ i are real and nonzero. If an eigenvalue λ \lambda λ had multiplicity at least six, the nonzero quadric h − λ q h-\lambda q h − λ q would have rank at most two, hence would factor over C \mathbb{C} C into two linear forms, possibly equal; as ( h , f ) = ( h , f A ) (h,f)=(h,f_A) ( h , f ) = ( h , f A ) is the prime ideal of the affine cone of A A A (§ 21 ), one factor would lie in it, contradicting full span. By the Jacobian criterion the projective singular locus is the union of the sets P ( W λ ) ∩ V ( q ) \mathbb{P}(W_\lambda)\cap V(q) P ( W λ ) ∩ V ( q ) over the relative eigenspaces W λ W_\lambda W λ ; an eigenspace of dimension m ≥ 2 m \ge2 m ≥ 2 contributes dimension m − 2 m-2 m − 2 , and one of dimension one contributes nothing. So the projective singular locus has dimension at most three, and that of the six-dimensional affine cone, vertex included, at most four. Both spaces are therefore regular in codimension one; as integral complete intersections they are Cohen–Macaulay, hence satisfy Serre’s second condition, and Serre’s normality criterion applies.
Proposition 23.2. If every relative eigenspace has dimension at most three, the standard graded cone ring S = C [ a 1 , … , a 8 ] / ( h , q ) S=\mathbb{C}[a_1,\ldots,a_8]/(h,q) S = C [ a 1 , … , a 8 ] / ( h , q ) is a unique factorization domain.
Proof. The calculation above now bounds the singular locus of the affine cone by dimension two. As dim S = 6 \dim S=6 dim S = 6 , every localization at a prime of height at most three is regular, hence factorial [29 Ch. 2, Theorem 1.6, p. 40] . The local complete-intersection factoriality theorem (Grothendieck’s theorem, in the form [29 §8, Remark 3 following Theorem 8.2, p. 27] ) applies to S m S_m S m , m = S > 0 m=S_{>0} m = S > 0 the vertex maximal ideal: its localizations at primes of height at most three are factorial. Finally, for the divisor class groups, Cl ( S ) → Cl ( S m ) \operatorname{Cl}(S)\to\operatorname{Cl}(S_m) Cl ( S ) → Cl ( S m ) is an isomorphism for this positively graded normal domain [29 Proposition 7.4, p. 21] ; the target is zero, so the normal Noetherian domain S S S is factorial.
Proposition 23.3. In the factorial case of Proposition 23.2, suppose the quadratic transpose Gram G = B ( a ) B ( a ) T G=B(a)B(a)^{T} G = B ( a ) B ( a ) T of Lemma 17.4 has generic rank two on A A A (so that, by Lemma 12.7, its generic
transpose kernel is a totally isotropic three-space for the standard eight-form). Then, as an ambient real quadratic matrix, G = h D 0 + q D 1 + u u T − v v T G = hD_0 + qD_1 + uu^{T} - vv^{T} G = h D 0 + q D 1 + u u T − v v T (balanced) or G = h D 0 + q D 1 + c b T + b c T G = hD_0 + qD_1 + cb^{T} + bc^{T} G = h D 0 + q D 1 + c b T + b c T with c ≠ 0 c \ne0 c = 0 (unbalanced), where D 0 , D 1 D_0, D_1 D 0 , D 1 are constant real symmetric matrices and the columns have the degrees of Lemma 12.8.
Proof. By Lemma 12.7, applied to A A A and to its affine cone, the pairing ⟨ G x , G x ′ ⟩ = x T G x ′ \langle Gx, Gx'\rangle= x^{T}Gx' ⟨ G x , G x ′ ⟩ = x T G x ′ induced by G G G on its image is hyperbolic over the real function field, and its two isotropic lines are individually real-defined.
The saturated modules of the two isotropic lines are rank-one reflexive modules over the normal domain S S S (the intersections of their height-one localizations inside their fraction-field lines). Factoriality makes their divisor classes zero, so they are principal. As S S S is normal, so is its conjugation-fixed subring S R S_{\mathbb{R}} S R , and Lemma 12.8 applies. Lift its homogeneous columns to ambient real polynomials; the difference between G G G and the lift has degree two and lies in ( h , q ) (h,q) ( h , q ) , so it is h D 0 + q D 1 hD_0 + qD_1 h D 0 + q D 1 with constant real symmetric D 0 , D 1 D_0,D_1 D 0 , D 1 .
For the full eight-coefficient pencil, Proposition 22.3, on the eight-dimensional coefficient space, applies to the unbalanced form, and Proposition 22.1 applies to the balanced form, since relative multiplicities at most three imply λ 2 < λ 7 \lambda_2 < \lambda_7 λ 2 < λ 7 . Thus Im B ( a ) = E X \operatorname{Im} B(a) = E_X Im B ( a ) = E X for every real a ≠ 0 a \ne0 a = 0 at every anchor of this case; on a real open set of such anchors Lemmas 16.2 and 16.1 and Proposition 9.1(2) make R R R Clifford (Section 27).
Extraction at the singular integral residue images
In the situation of Proposition 21.3(1), let q q q be the positive definite form of Proposition 23.1, and suppose that h h h has a relative eigenvalue λ \lambda λ with respect to q q q of multiplicity m ∈ { 4 , 5 } m \in\{4,5\} m ∈ { 4 , 5 } . Let W W W be its eigenspace and Z Z Z the q q q -orthogonal complement, so that dim W = m \dim W = m dim W = m and dim Z = 8 − m \dim Z = 8-m dim Z = 8 − m ; write a = ( w , z ) a = (w,z) a = ( w , z ) and q ( w , z ) = q W ( w ) + q Z ( z ) q(w,z) = q_W(w) + q_Z(z) q ( w , z ) = q W ( w ) + q Z ( z ) , and put ρ = h − λ q \rho= h-\lambda q ρ = h − λ q , a form in z z z alone that is nondegenerate on Z Z Z . The residue image is the real geometrically integral variety A = V ( h , q ) = V ( q , ρ ) ⊂ P 7 A = V(h,q) = V(q,\rho) \subset\mathbb{P}^{7} A = V ( h , q ) = V ( q , ρ ) ⊂ P 7 , and Proposition 23.1 gives normality of the standard graded complex cone ring S = C [ w , z ] / ( q W + q Z , ρ ) S = \mathbb{C}[w,z]/(q_W + q_Z,\rho) S = C [ w , z ] / ( q W + q Z , ρ ) . Let G G G be the real symmetric quadratic transpose Gram matrix. For type ( α , s ) = ( 5 , 5 ) (\alpha,s) = (5,5) ( α , s ) = ( 5 , 5 ) , Lemmas 17.4 and 10.3 give it rank zero or two on A A A and generic rank two. Hence, by Lemma 12.7, its generic transpose kernel is a totally isotropic three-space in the standard eight-form, and the binary pairing induced by G G G on its image is hyperbolic over the real function field, with its two isotropic lines individually real-defined. The singular locus of A A A contains the stratum { z = 0 , q W ( w ) = 0 } \{z = 0,\, q_W(w) = 0\} { z = 0 , q W ( w ) = 0 } , and, when the relative spectrum is ( 4 , 4 ) (4,4) ( 4 , 4 ) , also the stratum { w = 0 , q Z ( z ) = 0 } \{w = 0,\, q_Z(z) = 0\} { w = 0 , q Z ( z ) = 0 } ; these are the singular strata used below (Remark 24.7). If G G G vanishes on one of these strata, Proposition 22.4 applies; we extract a real normal form in the remaining case, where G G G has rank two at the generic points of those strata.
We use divisor classes of affine normal rings. Localization of a normal domain induces a surjection on divisor class groups whose kernel is generated by the height-one primes removed [^29, Theorem 6.3 and formula (6.4), p. 17]; its factoriality corollary is Nagata’s criterion [^29, Corollary to Theorem 6.3’, p. 18].
Lemma 24.1. Let F F F be a field of characteristic zero, N = F [ z 1 , z 2 , z 3 , z 4 ] / ( z 1 z 2 − z 3 z 4 ) \mathcal{N} = F[z_1,z_2,z_3,z_4]/(z_1z_2-z_3z_4) N = F [ z 1 , z 2 , z 3 , z 4 ] / ( z 1 z 2 − z 3 z 4 ) , m = ( z 1 , z 2 , z 3 , z 4 ) \mathfrak{m} = (z_1,z_2,z_3,z_4) m = ( z 1 , z 2 , z 3 , z 4 ) and P = ( z 1 , z 3 ) P = (z_1,z_3) P = ( z 1 , z 3 ) . The ring N ≅ F [ s u , s v , t u , t v ] \mathcal{N} \cong F[su,sv,tu,tv] N ≅ F [ s u , s v , t u , t v ] is a normal domain (a domain of rank four), so its divisor classes are defined, and the class of P N m P\mathcal{N}_{\mathfrak{m}} P N m has infinite order in Cl ( N m ) \operatorname{Cl}(\mathcal{N}_{\mathfrak{m}}) Cl ( N m ) .
Proof. In the Segre presentation z 1 = s u z_1 = su z 1 = s u , z 3 = s v z_3 = sv z 3 = s v , z 4 = t u z_4 = tu z 4 = t u , z 2 = t v z_2 = tv z 2 = t v , the valuation at P P P is the exponent of s s s . A homogeneous element of standard degree d d d has bidegree ( d , d ) (d,d) ( d , d ) in ( s , t ) (s,t) ( s , t ) and ( u , v ) (u,v) ( u , v ) , so its s s s -order is at most d d d . Hence the symbolic power P ( n ) P^{(n)} P ( n ) has no nonzero homogeneous element of degree less than n n n , and its degree-n n n part contains the n + 1 n+1 n + 1 independent monomials z 1 n − i z 3 i z_1^{n-i}z_3^i z 1 n − i z 3 i , 0 ≤ i ≤ n 0 \le i \le n 0 ≤ i ≤ n . Localization at m \mathfrak m m only introduces denominators with nonzero constant term, so it preserves initial degree, and every element of m P ( n ) N m \mathfrak mP^{(n)}N_{\mathfrak m} m P ( n ) N m has initial degree at least n + 1 n+1 n + 1 . So these monomials remain independent modulo m P ( n ) N m \mathfrak mP^{(n)}N_{\mathfrak m} m P ( n ) N m , and P ( n ) N m P^{(n)}N_{\mathfrak m} P ( n ) N m needs at least n + 1 n+1 n + 1 generators; it is not principal for any n ≥ 1 n \ge1 n ≥ 1 . In a normal domain, n [ P ] = 0 n[P] = 0 n [ P ] = 0 would make this divisorial symbolic power principal. Hence [ P ] [P] [ P ] has infinite order.
Lemma 24.2. For m = 5 m=5 m = 5 , let η \eta η , the singular-stratum prime, be the prime of S S S generated by the coordinates of z z z . Then Cl ( S ) ≅ Z / 2 \operatorname{Cl}(S) \cong\mathbb{Z}/2 Cl ( S ) ≅ Z /2 , and Cl ( S ) → Cl ( S η ) \operatorname{Cl}(S) \to\operatorname{Cl}(S_\eta) Cl ( S ) → Cl ( S η ) is injective.
Proof. Choose complex coordinates with ρ = z 1 z 2 − z 3 2 \rho=z_1z_2-z_3^2 ρ = z 1 z 2 − z 3 2 and q W = w 1 w 2 + w 3 2 + w 4 2 + w 5 2 q_W=w_1w_2+w_3^2+w_4^2+w_5^2 q W = w 1 w 2 + w 3 2 + w 4 2 + w 5 2 . Inverting z 1 z_1 z 1 and eliminating z 2 z_2 z 2 gives S [ z 1 − 1 ] = R [ w 1 , … , w 5 ] / ( w 1 w 2 + g ) S[z_1^{-1}]=R[w_1,\ldots,w_5]/(w_1w_2+g) S [ z 1 − 1 ] = R [ w 1 , … , w 5 ] / ( w 1 w 2 + g ) with R = C [ z 1 , z 1 − 1 , z 3 ] R=\mathbb{C}[z_1,z_1^{-1},z_3] R = C [ z 1 , z 1 − 1 , z 3 ] and g = w 3 2 + w 4 2 + w 5 2 + q Z ( z 1 , z 3 2 / z 1 , z 3 ) g=w_3^2+w_4^2+w_5^2+q_Z(z_1,z_3^2/z_1,z_3) g = w 3 2 + w 4 2 + w 5 2 + q Z ( z 1 , z 3 2 / z 1 , z 3 ) . Over Frac ( R ) \operatorname{Frac}(R) Frac ( R ) the quadratic leading part of g g g in the three w w w variables has rank three, whereas a product of affine linear factors would have leading rank at most two; so g g g is irreducible, and it is primitive over the factorial ring R R R . Thus w 1 w_1 w 1 is prime in S [ z 1 − 1 ] S[z_1^{-1}] S [ z 1 − 1 ] . Inverting w 1 w_1 w 1 also eliminates w 2 w_2 w 2 and gives a Laurent polynomial ring, so S [ z 1 − 1 ] S[z_1^{-1}] S [ z 1 − 1 ] is factorial by Nagata’s criterion.
The only height-one prime above z 1 z_1 z 1 is P = ( z 1 , z 3 ) P=(z_1,z_3) P = ( z 1 , z 3 ) : the relation modulo z 1 z_1 z 1 is z 3 2 = 0 z_3^2=0 z 3 2 = 0 , and the reduced quotient has a rank-five w w w -quadratic part and is a domain. At P P P , z 2 z_2 z 2 is a unit and z 1 = z 3 2 / z 2 z_1=z_3^2/z_2 z 1 = z 3 2 / z 2 . Hence div ( z 1 ) = 2 P \operatorname{div}(z_1)=2P div ( z 1 ) = 2 P , and localization shows that Cl ( S ) \operatorname{Cl}(S) Cl ( S ) is generated by [ P ] [P] [ P ] with 2 [ P ] = 0 2[P]=0 2 [ P ] = 0 .
At η \eta η , w 2 w_2 w 2 is invertible, and eliminating w 1 w_1 w 1 identifies S η S_\eta S η exactly with the vertex local ring of z 1 z 2 = z 3 2 z_1z_2=z_3^2 z 1 z 2 = z 3 2 over F = C ( w 2 , w 3 , w 4 , w 5 ) F=\mathbb{C}(w_2,w_3,w_4,w_5) F = C ( w 2 , w 3 , w 4 , w 5 ) . The three z z z variables are independent in its cotangent space, where ( z 1 , z 3 ) (z_1,z_3) ( z 1 , z 3 ) has a two-dimensional image, so ( z 1 , z 3 ) (z_1,z_3) ( z 1 , z 3 ) is not principal. Being a height-one prime of a normal local domain, it is divisorial, so its class is nonzero locally. A cyclic group of order dividing two whose generator has nonzero image maps injectively. This proves both assertions.
Lemma 24.3. Suppose m = 4 m=4 m = 4 and q Z q_Z q Z is not proportional to ρ \rho ρ . For the prime η \eta η of S S S generated by the coordinates of z z z , the map Cl ( S ) → Cl ( S η ) \operatorname{Cl}(S) \to\operatorname{Cl}(S_\eta) Cl ( S ) → Cl ( S η ) is injective, and Cl ( S ) ≅ Z \operatorname{Cl}(S) \cong\mathbb{Z} Cl ( S ) ≅ Z .
Proof. In complex coordinates with ρ = z 1 z 2 − z 3 z 4 \rho=z_1z_2-z_3z_4 ρ = z 1 z 2 − z 3 z 4 and q W = w 1 w 2 + w 3 w 4 q_W=w_1w_2+w_3w_4 q W = w 1 w 2 + w 3 w 4 , inverting z 1 z_1 z 1 gives S [ z 1 − 1 ] = R [ w 1 , … , w 4 ] / ( w 1 w 2 + w 3 w 4 + ψ ) S[z_1^{-1}]=R[w_1,\ldots,w_4]/(w_1w_2+w_3w_4+\psi) S [ z 1 − 1 ] = R [ w 1 , … , w 4 ] / ( w 1 w 2 + w 3 w 4 + ψ ) with R = C [ z 1 , z 1 − 1 , z 3 , z 4 ] R=\mathbb{C}[z_1,z_1^{-1},z_3,z_4] R = C [ z 1 , z 1 − 1 , z 3 , z 4 ] and ψ = q Z ( z 1 , z 3 z 4 / z 1 , z 3 , z 4 ) \psi=q_Z(z_1,z_3z_4/z_1,z_3,z_4) ψ = q Z ( z 1 , z 3 z 4 / z 1 , z 3 , z 4 ) . Our hypothesis gives ψ ≠ 0 \psi\ne0 ψ = 0 : otherwise q Z q_Z q Z would vanish on the irreducible ρ \rho ρ -quadric and, being quadratic, would be proportional to ρ \rho ρ . The polynomial w 3 w 4 + ψ w_3w_4+\psi w 3 w 4 + ψ is irreducible over Frac ( R ) \operatorname{Frac}(R) Frac ( R ) and primitive over R R R , so w 1 w_1 w 1 is prime; inverting it produces a Laurent polynomial ring, and as before S [ z 1 − 1 ] S[z_1^{-1}] S [ z 1 − 1 ] is factorial.
The two height-one primes above z 1 z_1 z 1 are P = ( z 1 , z 3 ) P=(z_1,z_3) P = ( z 1 , z 3 ) and P ′ = ( z 1 , z 4 ) P'=(z_1,z_4) P ′ = ( z 1 , z 4 ) ; their quotients are domains because the w w w -quadratic has rank four. The equation z 1 z 2 = z 3 z 4 z_1z_2=z_3z_4 z 1 z 2 = z 3 z 4 gives z 1 z_1 z 1 valuation one at each, so div ( z 1 ) = P + P ′ \operatorname{div}(z_1)=P+P' div ( z 1 ) = P + P ′ , and consequently Cl ( S ) \operatorname{Cl}(S) Cl ( S ) is cyclic, generated by [ P ] [P] [ P ] . At η \eta η , invert w 2 w_2 w 2 and eliminate w 1 = − ( w 3 w 4 + q Z ( z ) ) / w 2 w_1=-(w_3w_4+q_Z(z))/w_2 w 1 = − ( w 3 w 4 + q Z ( z )) / w 2 ; the local ring is exactly ( F [ z 1 , z 2 , z 3 , z 4 ] / ( z 1 z 2 − z 3 z 4 ) ) ( z 1 , z 2 , z 3 , z 4 ) (F[z_1,z_2,z_3,z_4]/(z_1z_2-z_3z_4))_{(z_1,z_2,z_3,z_4)} ( F [ z 1 , z 2 , z 3 , z 4 ] / ( z 1 z 2 − z 3 z 4 ) ) ( z 1 , z 2 , z 3 , z 4 ) with F = C ( w 2 , w 3 , w 4 ) F=\mathbb{C}(w_2,w_3,w_4) F = C ( w 2 , w 3 , w 4 ) . There P P P maps to the plane prime of Lemma 24.1, of infinite order, so the map from the cyclic global class group is injective, which proves the claim.
When q Z q_Z q Z is proportional to ρ \rho ρ , the relative spectrum has two four-dimensional eigenspaces, and the two defining equations give a product of two rank-four quadric cones.
Lemma 24.4. Let S = C [ w 1 , … , w 4 , z 1 , … , z 4 ] / ( w 1 w 2 − w 3 w 4 , z 1 z 2 − z 3 z 4 ) S=\mathbb{C}[w_1,\ldots,w_4,z_1,\ldots,z_4]/(w_1w_2-w_3w_4,\ z_1z_2-z_3z_4) S = C [ w 1 , … , w 4 , z 1 , … , z 4 ] / ( w 1 w 2 − w 3 w 4 , z 1 z 2 − z 3 z 4 ) , η W = ( w 1 , … , w 4 ) \eta_W=(w_1,\ldots,w_4) η W = ( w 1 , … , w 4 ) and η Z = ( z 1 , … , z 4 ) \eta_Z=(z_1,\ldots,z_4) η Z = ( z 1 , … , z 4 ) . Then Cl ( S ) ≅ Z 2 \operatorname{Cl}(S)\cong\mathbb{Z}^2 Cl ( S ) ≅ Z 2 , and Cl ( S ) → Cl ( S η W ) ⊕ Cl ( S η Z ) \operatorname{Cl}(S)\to\operatorname{Cl}(S_{\eta_W})\oplus\operatorname{Cl}(S_{\eta_Z}) Cl ( S ) → Cl ( S η W ) ⊕ Cl ( S η Z ) is injective.
Proof. Inverting w 1 w_1 w 1 and z 1 z_1 z 1 eliminates w 2 w_2 w 2 , z 2 z_2 z 2 and gives a Laurent polynomial ring. The removed height-one primes are P W = ( w 1 , w 3 ) P_W=(w_1,w_3) P W = ( w 1 , w 3 ) , P W ′ = ( w 1 , w 4 ) P'_W=(w_1,w_4) P W ′ = ( w 1 , w 4 ) , P Z = ( z 1 , z 3 ) P_Z=(z_1,z_3) P Z = ( z 1 , z 3 ) and P Z ′ = ( z 1 , z 4 ) P'_Z=(z_1,z_4) P Z ′ = ( z 1 , z 4 ) , and the principal divisors of w 1 , z 1 w_1,z_1 w 1 , z 1 give [ P W ′ ] = − [ P W ] [P'_W]=-[P_W] [ P W ′ ] = − [ P W ] and [ P Z ′ ] = − [ P Z ] [P'_Z]=-[P_Z] [ P Z ′ ] = − [ P Z ] ; so [ P W ] , [ P Z ] [P_W],[P_Z] [ P W ] , [ P Z ] generate the class group. At η W \eta_W η W the local ring is that of the cone w 1 w 2 = w 3 w 4 w_1w_2=w_3w_4 w 1 w 2 = w 3 w 4 at its vertex, over the function field of the z z z -quadric; the class of P W P_W P W has infinite order by Lemma 24.1, while P Z P_Z P Z becomes the unit ideal since z 1 ∉ η W z_1\notin\eta_W z 1 ∈ / η W . At η Z \eta_Z η Z the roles reverse. So n [ P W ] + k [ P Z ] n[P_W]+k[P_Z] n [ P W ] + k [ P Z ] vanishes in both local groups only if n = k = 0 n=k=0 n = k = 0 , which proves injectivity and excludes every relation between the two global generators.
Lemma 24.5 (Local splitting). Let T T T be a normal local domain in which 2 2 2 is invertible. A unimodular symmetric bilinear form on a free T T T -module of rank two which is hyperbolic over Frac ( T ) \operatorname{Frac}(T) Frac ( T ) splits over T T T into two free isotropic lines.
Proof. There is a vector whose norm is a unit of T T T : otherwise all diagonal entries and, by evaluation on sums of basis vectors, all matrix entries would lie in the maximal ideal. Splitting it off gives diag ( a , b ) \operatorname{diag}(a,b) diag ( a , b ) with units a , b a,b a , b . Generic hyperbolicity gives r 2 = − b / a r^2=-b/a r 2 = − b / a in the fraction field; r r r is integral over T T T , hence in T T T by normality, and a unit. The isotropic vectors ( r , 1 ) , ( − r , 1 ) (r,1),(-r,1) ( r , 1 ) , ( − r , 1 ) form a basis, since 2 r 2r 2 r is a unit.
Proposition 24.6. In the setup of this section, suppose G G G has rank two at the generic singular-stratum prime in Lemma 24.2 or Lemma 24.3, or at both primes in Lemma 24.4, as applicable. Then, in ambient real coefficient coordinates, G = q D 0 + ρ D 1 + u u T − v v T G=qD_0+\rho D_1+uu^{T}-vv^{T} G = q D 0 + ρ D 1 + u u T − v v T (balanced) or G = q D 0 + ρ D 1 + c b T + b c T G=qD_0+\rho D_1+cb^{T}+bc^{T} G = q D 0 + ρ D 1 + c b T + b c T (unbalanced), where D 0 , D 1 D_0,D_1 D 0 , D 1 are constant real symmetric matrices, u , v u,v u , v are real linear columns, c c c is a nonzero constant real column, and b b b is a real quadratic column. No primitivity of G G G is assumed: common divisorial factors are allowed.
Proof. Over Frac ( S ) \operatorname{Frac}(S) Frac ( S ) the hyperbolic binary pairing defines two isotropic image lines of G G G . For each such line L L L let L = L ∩ S N ⊂ Frac ( S ) N \mathcal{L}=L\cap S^N\subset\operatorname{Frac}(S)^N L = L ∩ S N ⊂ Frac ( S ) N be its saturated closure, N N N the matrix size. It is a rank-one reflexive module: its height-one localizations are saturated line lattices, and normality identifies their intersection with L \mathcal{L} L .
At each prime η \eta η of Lemmas 24.2–24.4 at which G G G has rank two by hypothesis, a unit two-minor and vanishing three-minors of G G G make its image a free rank-two direct summand, and the induced symmetric pairing is unimodular: viewing G G G as a symmetric map from the dual free module, it is an isomorphism onto its image after quotienting its kernel, and its inverse defines the pairing. By the isometry K 0 ⊥ / K 0 → Im G K_0^\perp/K_0\to\operatorname{Im}G K 0 ⊥ / K 0 → Im G induced by B B B (Lemma 12.7), this is the pairing whose generic isotropic lines were used above. Lemma 24.5 makes each line closure free at η \eta η . The injectivity in Lemmas 24.2–24.4 therefore makes each global class zero; in the case of Lemma 24.4 this uses both primes η W \eta_W η W and η Z \eta_Z η Z .
A class-zero rank-one reflexive module is free, so both closures are principal. As S S S is normal, so is its conjugation-fixed subring S R = R [ w , z ] / ( q , ρ ) S_{\mathbb{R}}=\mathbb{R}[w,z]/(q,\rho) S R = R [ w , z ] / ( q , ρ ) , and the pairing is hyperbolic over Frac ( S R ) \operatorname{Frac}(S_{\mathbb{R}}) Frac ( S R ) ; so Lemma 12.8 applies and gives, modulo the defining ideal, the balanced form with real linear columns u , v u,v u , v or the unbalanced form with a nonzero constant real column c c c and a real quadratic column b b b . Real homogeneous elements lift to real ambient polynomials, and the degree-two part of the ideal of A A A is precisely span R { q , ρ } \operatorname{span}_{\mathbb{R}}\{q,\rho\} span R { q , ρ } , so the remainders are q D 0 + ρ D 1 qD_0+\rho D_1 q D 0 + ρ D 1 with real symmetric D 0 , D 1 D_0,D_1 D 0 , D 1 .
Remark 24.7. For a single multiplicity-four or multiplicity-five block the relevant stratum is z = 0 z=0 z = 0 , q W ( w ) = 0 q_W(w)=0 q W ( w ) = 0 . For relative spectrum ( 4 , 4 ) (4,4) ( 4 , 4 ) there are two strata, one for each cone vertex; for the stratum w = 0 w=0 w = 0 , q Z ( z ) = 0 q_Z(z)=0 q Z ( z ) = 0 , apply Proposition 22.4 with W W W and Z Z Z exchanged, with h = λ ′ q + ρ ′ h=\lambda' q+\rho' h = λ ′ q + ρ ′ , where λ ′ \lambda' λ ′ is the relative eigenvalue on Z Z Z and ρ ′ = ( λ − λ ′ ) q W \rho'=(\lambda-\lambda')q_W ρ ′ = ( λ − λ ′ ) q W is nondegenerate on W W W . If the Gram vanishes identically on a required stratum, Proposition 22.4 applies. Otherwise its pointwise ranks zero or two give rank two at every required generic point, and Proposition 24.6 applies. The balanced and unbalanced alignment results, Propositions 22.2 and 22.3, then give Im B ( a ) = E X \operatorname{Im} B(a)=E_X Im B ( a ) = E X for every real a ≠ 0 a\ne0 a = 0 at this anchor, and § 27 concludes (Lemmas 16.2 and 16.1 on a real open set of anchors).
The conjugate five-dimensional images: first reductions
We work in Setting (T) of § 12 , with the transpose pencil B ( a ) B(a) B ( a ) at a regular anchor with ν = 8 \nu=8 ν = 8 and G = B B T G=BB^{\mathsf{T}} G = B B T . Proposition 21.5 gives real coordinates a = ( x , z ) ∈ W ⊕ Z a=(x,z)\in W\oplus Z a = ( x , z ) ∈ W ⊕ Z , dim W = 6 \dim W=6 dim W = 6 , dim Z = 2 \dim Z=2 dim Z = 2 , with
q ( x , z ) = ∣ x ∣ 2 + ∣ z ∣ 2 , h ( x , z ) = λ q ( x , z ) + ρ ( z ) , (58) q(x,z)=|x|^2+|z|^2,\qquad h(x,z)=\lambda q(x,z)+\rho(z), \tag*{(58)} q ( x , z ) = ∣ x ∣ 2 + ∣ z ∣ 2 , h ( x , z ) = λ q ( x , z ) + ρ ( z ) , (58)
where the definite binary form ρ \rho ρ is a nonzero real multiple of ℓ ℓ ˉ \ell\bar{\ell} ℓ ℓ ˉ for the complex linear form ℓ \ell ℓ of Proposition 21.3(2). Over C \mathbb{C} C the two null lines of ρ \rho ρ give distinct conjugate components A + = V ( h , ℓ ) A_+=V(h,\ell) A + = V ( h , ℓ ) , A − = V ( h , ℓ ˉ ) A_-=V(h,\bar{\ell}) A − = V ( h , ℓ ˉ ) , and A + ∪ A − = V ( h , q ) = V ( q , ρ ) A_+\cup A_-=V(h,q)=V(q,\rho) A + ∪ A − = V ( h , q ) = V ( q , ρ ) is a reduced complete intersection with ideal ( q , ρ ) = ( h , q ) (q,\rho)=(h,q) ( q , ρ ) = ( h , q ) : ( q , ℓ ) (q,\ell) ( q , ℓ ) and ( q , ℓ ˉ ) (q,\bar{\ell}) ( q , ℓ ˉ ) are prime, and in the factorial ring C [ a ] / ( q ) \mathbb{C}[a]/(q) C [ a ] / ( q ) [29 Theorem 8.2, p. 26] the nonassociate primes ℓ , ℓ ˉ \ell,\bar{\ell} ℓ , ℓ ˉ satisfy ( ℓ ) ∩ ( ℓ ˉ ) = ( ℓ ℓ ˉ ) (\ell)\cap(\bar{\ell})=(\ell\bar{\ell}) ( ℓ ) ∩ ( ℓ ˉ ) = ( ℓ ℓ ˉ ) . By Proposition 21.3, Remark 21.6 and Lemmas 17.4, 17.6 and 10.3 in type ( α , s ) = ( 5 , 5 ) (\alpha,s)=(5,5) ( α , s ) = ( 5 , 5 ) , each component spans its nonreal hyperplane and is a quadric of rank six or seven, and the Gram has rank at most two at every point of the union, generically two on each component; by Lemma 10.3 its rank is even at every point. On complex vectors ∣ ⋅ ∣ 2 |\cdot|^2 ∣ ⋅ ∣ 2 denotes the complex-bilinear extension of the real form. Set Σ = A + ∩ A − = { z = 0 , ∣ x ∣ 2 = 0 } \Sigma=A_+\cap A_-=\{z=0,\ |x|^2=0\} Σ = A + ∩ A − = { z = 0 , ∣ x ∣ 2 = 0 } .
Proposition 25.1. If G G G vanishes identically on Σ \Sigma Σ , then Im B ( a ) = E \operatorname{Im} B(a)=E Im B ( a ) = E for every nonzero real a a a .
Proof. Lemma 12.9 applies with dim W = 6 \dim W=6 dim W = 6 : its hypothesis (i) is our assumption, and (ii) holds because the Gram has rank at most two at every point of V ( q , ρ ) = A + ∪ A − V(q,\rho)=A_+\cup A_- V ( q , ρ ) = A + ∪ A − . As dim Z = 2 \dim Z=2 dim Z = 2 and ρ \rho ρ is definite, hence nondegenerate, part (3) of that lemma gives the claim. □ \square □
Proposition 25.2. Suppose that on A + A_+ A + one has G = u b T + b u T G=ub^{\mathsf{T}}+bu^{\mathsf{T}} G = u b T + b u T , where u = u 1 + i u 2 u=u_1+iu_2 u = u 1 + i u 2 , with u 1 , u 2 ∈ V u_1,u_2\in V u 1 , u 2 ∈ V , is a fixed nonzero complex vector and b b b is a homogeneous quadratic vector in the component coordinate ring. Then Im B ( a ) = E \operatorname{Im} B(a)=E Im B ( a ) = E for every nonzero real a a a .
Proof. Let F = span R { u 1 , u 2 } \mathcal{F}=\operatorname{span}_{\mathbb{R}}\{u_1,u_2\} F = span R { u 1 , u 2 } , so dim F ≤ 2 \dim\mathcal{F}\le2 dim F ≤ 2 , and let P P P be real orthogonal projection to F ⊥ \mathcal{F}^{\perp} F ⊥ . Then P G P PGP P GP vanishes on A + A_+ A + , and by reality on A − A_- A − . As the reduced union has degree-two ideal span { h , q } \{h,q\} { h , q } , P G P = h D 0 + q D 1 PGP=hD_0+qD_1 P GP = h D 0 + q D 1 with constant real symmetric D 0 , D 1 D_0,D_1 D 0 , D 1 ; it is positive semidefinite of rank at most eight at every real point. On q = 1 q=1 q = 1 the Rayleigh range of h h h is a nontrivial interval [ t 0 , t 1 ] [t_0,t_1] [ t 0 , t 1 ] , and every D 1 + t D 0 D_1+tD_0 D 1 + t D 0 with t ∈ [ t 0 , t 1 ] t\in[t_0,t_1] t ∈ [ t 0 , t 1 ] is positive semidefinite, so by Lemma 12.1 all of them with interior t t t have one image V 0 V_0 V 0 , of dimension at most eight. On the nonempty real open set of interior Rayleigh values, ( P B ) ( P B ) T (PB)(PB)^{\mathsf{T}} ( P B ) ( P B ) T has image V 0 V_0 V 0 , so the projection of the linear pencil P B PB P B to V 0 ⊥ V_0^\perp V 0 ⊥ vanishes there, hence identically. Thus all images of B B B lie in the fixed real space F + V 0 \mathcal{F}+V_0 F + V 0 of dimension at most ten, and Lemma 16.3 applies to the reduced union. For its component hypothesis, a nonzero real linear form vanishing on A + A_+ A + would vanish on its span ker ℓ \ker\ell ker ℓ , hence would be a complex multiple of ℓ \ell ℓ and, being real, also of ℓ ˉ \bar{\ell} ℓ ˉ , which is impossible as ℓ , ℓ ˉ \ell,\bar{\ell} ℓ , ℓ ˉ are independent; likewise for A − A_- A − . □ \square □
By Lemma 12.7, applied to A + A_+ A + , the rank-two symmetric pairing induced by G G G on its image is hyperbolic over the function field of A + A_+ A + .
Proposition 25.3. If Im B ( a ) ≠ E \operatorname{Im} B(a)\ne E Im B ( a ) = E for some nonzero real a a a , then G = p r T + r p T G=pr^{\mathsf{T}}+rp^{\mathsf{T}} G = p r T + r p T on A + A_+ A + with complex homogeneous linear columns p , r p,r p , r ; on A − A_- A − the factorization is their conjugate. Proof. The ring S + = C [ a 1 , … , a 8 ] / ( h , ℓ ) S_+ = \mathbb{C}[a_1,\ldots,a_8]/(h,\ell) S + = C [ a 1 , … , a 8 ] / ( h , ℓ ) is a UFD by Klein–Nagata [29 Theorem 8.2, p. 26] , as the restricted quadric has rank six or seven (with a polynomial extension if there is a radical variable). So the saturated rank-one reflexive modules of the two isotropic lines of the generic rank-two pairing are principal over the UFD, and Lemma 12.8 applies: G = t ( p r T + r p T ) G = t(pr^{T} + rp^{T}) G = t ( p r T + r p T ) on A + A_+ A + with homogeneous t ∈ S + t \in S_+ t ∈ S + and homogeneous columns p , r p,r p , r of nonnegative degrees, and deg t + deg p + deg r = 2 \deg t + \deg p + \deg r = 2 deg t + deg p + deg r = 2 . If a column has degree zero, absorb t t t and the other column into the quadratic vector of Proposition 25.2; that proposition then gives Im B ( a ) = E \operatorname{Im} B(a) = E Im B ( a ) = E for every nonzero real a a a , contrary to the hypothesis. Otherwise both columns are linear and t t t is a nonzero complex constant, which we absorb into one column.
Gluing the conjugate residue components
We keep Setting (T) of Section 12 and the notation of Section 25; the radial isomorphism L L L need not be an isometry.
Lemma 26.1. Let W ⊂ R 8 W \subset\mathbb{R}^8 W ⊂ R 8 have dimension six, and let q W q_W q W be positive definite on W W W . There do not exist constant real symmetric operators D , D ′ D,D' D , D ′ on V V V and a nonzero real quadratic form s s s on W W W such that
G ( x ) = q W ( x ) D + s ( x ) D ′ ( x ∈ W ) , (59) G(x) = q_W(x)D + s(x)D' \qquad(x \in W), \tag*{(59)} G ( x ) = q W ( x ) D + s ( x ) D ′ ( x ∈ W ) , (59) where rank s ≤ 2 \operatorname{rank}s \le2 rank s ≤ 2 and D ′ D' D ′ has rank two and signature ( 1 , 1 ) (1,1) ( 1 , 1 ) .
Proof. As s , q W s,q_W s , q W are independent, the range of s s s on the ellipsoid q W = 1 q_W = 1 q W = 1 is a nondegenerate closed interval, and every D + t D ′ D+tD' D + t D ′ on it is positive semidefinite of rank eight. By Lemma 12.1, equality of the ranks gives these matrices a common kernel, endpoints included, so B ( x ) B(x) B ( x ) has a fixed image F F F of dimension eight for every real x ∈ W ∖ 0 x \in W \setminus0 x ∈ W ∖ 0 , and Lemma 12.5(1) gives F = E F = E F = E .
By independence of q W , s q_W,s q W , s , the operators D , D ′ D,D' D , D ′ are supported on E E E . The radical of s s s has dimension at least four, and at a nonzero vector in it (26.1) shows that D ∣ E D|_E D ∣ E is positive definite. So the two nonzero eigenvalues of D − 1 / 2 D ′ D − 1 / 2 ∣ E D^{-1/2}D'D^{-1/2}|_E D − 1/2 D ′ D − 1/2 ∣ E are α > 0 > β \alpha> 0 > \beta α > 0 > β , and determinants on E E E give
det ( G ( x ) ∣ E ) = det ( D ∣ E ) q W ( x ) 6 ( q W ( x ) + α s ( x ) ) ( q W ( x ) + β s ( x ) ) . \det(G(x)|_E) = \det(D|_E)q_W(x)^6\bigl(q_W(x) + \alpha s(x)\bigr)\bigl(q_W(x) + \beta s(x)\bigr). det ( G ( x ) ∣ E ) = det ( D ∣ E ) q W ( x ) 6 ( q W ( x ) + α s ( x ) ) ( q W ( x ) + β s ( x ) ) .
The last two factors are positive on the radical of s s s , and neither vanishes on the connected set W ∖ 0 W \setminus0 W ∖ 0 , because their product is positive there; so they are positive definite, irreducible real quadratics of rank six. Independence of q W , s q_W,s q W , s and α ≠ β \alpha\ne\beta α = β make the three quadratic factors pairwise nonassociate, so the last two occur to odd multiplicity. But in orthonormal coordinates on E E E the matrix B ( x ) B(x) B ( x ) is square and det ( G ( x ) ∣ E ) = ( det B ( x ) ) 2 \det(G(x)|_E) = (\det B(x))^2 det ( G ( x ) ∣ E ) = ( det B ( x ) ) 2 , contradicting unique factorization in the real polynomial ring on W W W .
Keep the coordinates a = ( x , z ) ∈ W ⊕ Z a = (x,z) \in W \oplus Z a = ( x , z ) ∈ W ⊕ Z , the forms q , ρ , ℓ q,\rho,\ell q , ρ , ℓ and the sets A ± , Σ A_\pm,\Sigma A ± , Σ of Section 25, and put q W ( x ) = ∣ x ∣ 2 q_W(x)=|x|^2 q W ( x ) = ∣ x ∣ 2 , U + = ker ℓ U_+ = \ker\ell U + = ker ℓ , U − = ker ℓ U_- = \ker\ell U − = ker ℓ . As ρ \rho ρ vanishes on U ± U_\pm U ± , A ± = V ( q ) ∩ P ( U ± ) A_\pm= V(q) \cap\mathbb{P}(U_\pm) A ± = V ( q ) ∩ P ( U ± ) , and Σ = V ( q W ) ⊂ P ( W C ) \Sigma= V(q_W) \subset\mathbb{P}(W_\mathbb{C}) Σ = V ( q W ) ⊂ P ( W C ) . These quadrics have rank at least six, so their homogeneous coordinate rings are factorial [29 Theorem 8.2, p. 26] , also with the additional transverse variable on either component.
Proposition 26.2. Assume that on A + A_+ A + there are complex linear columns p , r p,r p , r with G = p r T + r p T G = pr^{T} + rp^{T} G = p r T + r p T , that G G G has generic rank two on Σ \Sigma Σ , and that its induced binary pairing there is hyperbolic over R ( Σ ) \mathbb{R}(\Sigma) R ( Σ ) . Then there exist real linear columns p ^ , r ^ \hat p,\hat r p ^ , r ^ on R 8 \mathbb{R}^8 R 8 and constant real symmetric operators D 0 , D 1 D_0,D_1 D 0 , D 1 such that
G = q D 0 + ρ D 1 + p ^ r ^ T + r ^ p ^ T . (60) G = qD_0 + \rho D_1 + \hat p\hat r^{T} + \hat r\hat p^{T}. \tag*{(60)} G = q D 0 + ρ D 1 + p ^ r ^ T + r ^ p ^ T . (60)
Proof. Write p 0 = p ∣ W p_0=p|_W p 0 = p ∣ W , r 0 = r ∣ W r_0=r|_W r 0 = r ∣ W . As G G G has rank two on Σ \Sigma Σ , they are nonzero and independent over C ( Σ ) \mathbb{C}(\Sigma) C ( Σ ) and span the two isotropic lines of the induced binary pairing; hyperbolicity over R ( Σ ) \mathbb{R}(\Sigma) R ( Σ ) makes each line invariant under conjugation.
Let S 0 = C [ W ] / ( q W ) S_0=\mathbb{C}[W]/(q_W) S 0 = C [ W ] / ( q W ) , and suppose first that p 0 p_0 p 0 is primitive over this factorial ring. Its conjugate is primitive and spans the same fraction-field line; the proportionality factor and its inverse have nonnegative valuation at every height-one prime, so the factor is a unit of S 0 S_0 S 0 , hence a nonzero constant, of modulus one by conjugating twice. Rescale p p p by a constant and r r r inversely so that p 0 p_0 p 0 is real. Reality of G G G on Σ \Sigma Σ gives p 0 ( r 0 − r ˉ 0 ) T + ( r 0 − r ˉ 0 ) p 0 T = 0 p_0(r_0-\bar r_0)^{\mathsf T}+(r_0-\bar r_0)p_0^{\mathsf T}=0 p 0 ( r 0 − r ˉ 0 ) T + ( r 0 − r ˉ 0 ) p 0 T = 0 , and v ↦ p 0 v T + v p 0 T v\mapsto p_0v^{\mathsf T}+vp_0^{\mathsf T} v ↦ p 0 v T + v p 0 T is injective (a nonzero coordinate of p 0 p_0 p 0 , its diagonal entry and its row force v = 0 v=0 v = 0 , as in the proof of Lemma 12.8); so r 0 = r ˉ 0 r_0=\bar r_0 r 0 = r ˉ 0 in S 0 S_0 S 0 . As the ideal ( q W ) (q_W) ( q W ) has no degree-one part, both restrictions are real linear columns on W W W . The argument is symmetric if r 0 r_0 r 0 is primitive.
If neither restriction is primitive, homogeneity gives p 0 = f c p_0=fc p 0 = f c and r 0 = g d r_0=gd r 0 = g d with nonzero complex linear forms f , g f,g f , g and independent constant vectors c , d c,d c , d , which conjugation invariance of each constant line lets us take real by rescaling f , g f,g f , g . Then D ′ = c d T + d c T D'=cd^{\mathsf T}+dc^{\mathsf T} D ′ = c d T + d c T is real symmetric of rank two and signature ( 1 , 1 ) (1,1) ( 1 , 1 ) . Reality of G G G on Σ \Sigma Σ implies f g − f g ‾ ∈ C q W fg-\overline{fg}\in\mathbb{C}q_W f g − f g ∈ C q W ; the left side has rank at most four and q W q_W q W has rank six, so s = f g s=fg s = f g is a real polynomial, nonzero and of rank at most two. Lifting the quadratic identity on Σ \Sigma Σ gives G ( x , 0 ) = q W ( x ) D + s ( x ) D ′ G(x,0)=q_W(x)D+s(x)D' G ( x , 0 ) = q W ( x ) D + s ( x ) D ′ with D D D constant real symmetric, contrary to Lemma 26.1 . So at least one restriction is primitive, and both are real.
Choose a nonzero ζ ∈ U + ∩ Z C \zeta\in U_+\cap Z_{\mathbb C} ζ ∈ U + ∩ Z C . As ζ , ζ ˉ \zeta,\bar\zeta ζ , ζ ˉ are independent and the prescriptions commute with conjugation, there is a unique real linear column p ^ \hat p p ^ with p ^ ∣ W = p 0 \hat p|_W=p_0 p ^ ∣ W = p 0 , p ^ ( ζ ) = p ( ζ ) \hat p(\zeta)=p(\zeta) p ^ ( ζ ) = p ( ζ ) and p ^ ( ζ ˉ ) = p ( ζ ˉ ) \hat p(\bar\zeta)=p(\bar\zeta) p ^ ( ζ ˉ ) = p ( ζ ˉ ) ; define r ^ \hat r r ^ similarly. These restrict to p , r p,r p , r on U + U_+ U + and to their conjugates on U − U_- U − , so G − p ^ r ^ T − r ^ p ^ T G-\hat p\hat r^{\mathsf T}-\hat r\hat p^{\mathsf T} G − p ^ r ^ T − r ^ p ^ T vanishes on both component quadrics. In the factorial ring C [ a ] / ( q ) \mathbb{C}[a]/(q) C [ a ] / ( q ) the distinct primes ℓ , ℓ ˉ \ell,\bar\ell ℓ , ℓ ˉ satisfy ( ℓ ) ∩ ( ℓ ˉ ) = ( ℓ ℓ ˉ ) (\ell)\cap(\bar\ell)=(\ell\bar\ell) ( ℓ ) ∩ ( ℓ ˉ ) = ( ℓ ℓ ˉ ) , so the reduced union has quadratic ideal span C { q , ρ } \operatorname{span}_{\mathbb C}\{q,\rho\} span C { q , ρ } , and reality and symmetry give (26.2) with real symmetric coefficients.
For the transpose pencil the first hypothesis is Proposition 25.3, and the last two hold unless G G G vanishes identically on Σ \Sigma Σ . By Lemma 10.3 , rank G = 2 rank B − 8 \operatorname{rank}G=2\operatorname{rank}B-8 rank G = 2 rank B − 8 is even and at most two at every point of Σ ⊂ A + \Sigma\subset A_+ Σ ⊂ A + . As Σ \Sigma Σ is irreducible, either G G G vanishes identically on Σ \Sigma Σ , the case of Proposition 25.1, or G G G has generic rank two on Σ \Sigma Σ . In the latter case Lemma 12.7, applied to the real-defined quadric Σ \Sigma Σ , makes the induced binary pairing hyperbolic over R ( Σ ) \mathbb{R}(\Sigma) R ( Σ ) .
Finally, u = ( p ^ + r ^ ) / 2 u=(\hat p+\hat r)/\sqrt{2} u = ( p ^ + r ^ ) / 2 and v = ( p ^ − r ^ ) / 2 v=(\hat p-\hat r)/\sqrt{2} v = ( p ^ − r ^ ) / 2 turn the last term of (26.2) into u u T − v v T uu^{\mathsf T}-vv^{\mathsf T} u u T − v v T . As span { q , ρ } = span { q , h } \operatorname{span}\{q,\rho\}=\operatorname{span}\{q,h\} span { q , ρ } = span { q , h } , this is the real balanced form required by Proposition 22.1 , whose Setting (T) survives the real change of coefficient coordinates to $(x,z) ({xref}Section 12). The sixfold relative eigenvalue is extreme in this conjugate configuration, so {xref}Proposition 22.2 supplies $\lambda_2<\lambda_7$, and {xref}Proposition 22.1` gives Im B ( a ) = E \operatorname{Im}B(a)=E Im B ( a ) = E for every nonzero real a a a .
Assembly of the proof of Theorem A
The anchor set and the pole alternative
Fix R R R as in Theorem A. If R R R has two reduced eigenvalues, Theorem C applies. Theorems C , E and F classify twenty of the thirty-six multiplicity patterns on R 16 \mathbb{R}^{16} R 16 that contain a multiplicity 7 or 8 (the two-eigenvalue pattern ( 8 , 7 ) (8,7) ( 8 , 7 ) and nineteen with three or more eigenvalues), leaving exactly sixteen:
( 8 , 4 , 3 ) (8,4,3) ( 8 , 4 , 3 ) , ( 8 , 3 , 3 , 1 ) (8,3,3,1) ( 8 , 3 , 3 , 1 ) , ( 8 , 3 , 2 , 2 ) (8,3,2,2) ( 8 , 3 , 2 , 2 ) , ( 8 , 3 , 2 , 1 , 1 ) (8,3,2,1,1) ( 8 , 3 , 2 , 1 , 1 ) , ( 8 , 3 , 1 4 ) (8,3,1^4) ( 8 , 3 , 1 4 ) ;
( 7 , 5 , 3 ) (7,5,3) ( 7 , 5 , 3 ) , ( 7 , 4 , 4 ) (7,4,4) ( 7 , 4 , 4 ) , ( 7 , 4 , 3 , 1 ) (7,4,3,1) ( 7 , 4 , 3 , 1 ) , ( 7 , 4 , 2 , 2 ) (7,4,2,2) ( 7 , 4 , 2 , 2 ) , ( 7 , 4 , 2 , 1 , 1 ) (7,4,2,1,1) ( 7 , 4 , 2 , 1 , 1 ) , ( 7 , 4 , 1 4 ) (7,4,1^4) ( 7 , 4 , 1 4 ) ,
( 7 , 3 , 3 , 2 ) (7,3,3,2) ( 7 , 3 , 3 , 2 ) , ( 7 , 3 , 3 , 1 , 1 ) (7,3,3,1,1) ( 7 , 3 , 3 , 1 , 1 ) , ( 7 , 3 , 2 , 2 , 1 ) (7,3,2,2,1) ( 7 , 3 , 2 , 2 , 1 ) , ( 7 , 3 , 2 , 1 3 ) (7,3,2,1^3) ( 7 , 3 , 2 , 1 3 ) , ( 7 , 3 , 1 5 ) (7,3,1^5) ( 7 , 3 , 1 5 ) .
Shift the eigenvalue of multiplicity 8 or 7 to zero. The shifted family has constant rank ν = 7 \nu= 7 ν = 7 or ν = 8 \nu= 8 ν = 8 on R 16 ∖ 0 \mathbb{R}^{16} \setminus0 R 16 ∖ 0 , and its nonzero spectral matrix Λ \Lambda Λ is nonscalar. If Λ \Lambda Λ is definite, Theorem G shows that R R R is Clifford; from now on Λ \Lambda Λ is indefinite.
The construction of Lemma 9.3 applies at an anchor x x x whenever the generic-pair conditions of [23 Section 4, Lemmas 2(1) and 3, pp. 7–9 of arXiv:math/0204258v2] hold for ( x , y ) (x,y) ( x , y ) with y y y in a dense subset of x ⊥ x^\perp x ⊥ ; the condition Im J x ∩ Im J y = 0 \operatorname{Im} J_x \cap\operatorname{Im} J_y = 0 Im J x ∩ Im J y = 0 is the nonvanishing of the sum of squares of the 2 ν 2\nu 2 ν -minors of the 16 × 32 16 \times32 16 × 32 matrix [ J x J y ] [J_x\ J_y] [ J x J y ] , that is, rank [ J x J y ] = 2 ν \operatorname{rank}[J_x\ J_y] = 2\nu rank [ J x J y ] = 2 ν , a polynomial condition g ( x , y ) ≠ 0 g(x,y) \ne0 g ( x , y ) = 0 . Substituting y = ( ∥ x ∥ 2 I − x x T ) z y = (\lVert x\rVert^2 I - xx^T)z y = (∥ x ∥ 2 I − x x T ) z and expanding in z z z , the anchors at which g ( x , ⋅ ) g(x,\cdot) g ( x , ⋅ ) vanishes identically form a real algebraic set, proper because a good generic pair exists. Remove it, together with the proper algebraic exceptional loci used in the preceding sections: the non-generic anchors of Lemma 11.1 and Propositions 11.3 and 11.4 (a semialgebraic set of dimension less than sixteen, hence inside a proper algebraic set) and the loci used for spreading out geometric factors (Remark 27.2 ). The result is a nonempty Euclidean-open cone U 0 ⊂ R 16 ∖ 0 U_0 \subset\mathbb{R}^{16} \setminus0 U 0 ⊂ R 16 ∖ 0 ; its points are the regular anchors of § 10 (the construction at x x x is understood after normalizing to x / ∥ x ∥ x/\lVert x\rVert x / ∥ x ∥ ). For x ∈ U 0 x \in U_0 x ∈ U 0 we write F x F_x F x for the rational upper block and P x P_x P x for the linear lower block, of size ( 16 − ν ) × ν (16-\nu) \times\nu ( 16 − ν ) × ν .
By Lemma 10.1 , if F x F_x F x has no pole for one x ∈ U 0 x \in U_0 x ∈ U 0 , then R R R is Clifford. So we may assume that F x F_x F x has a pole for every x ∈ U 0 x \in U_0 x ∈ U 0 ; the rest of the proof shows that this again leads to a Clifford tensor.
A finite semialgebraic case partition
At each x ∈ U 0 x \in U_0 x ∈ U 0 the pole data of F x F_x F x consist of finitely many discrete invariants: the real factorization type of the reduced denominator and its pole orders; the ranks of the skew residues; the real residue support d d d ; the residue-variety dimension α \alpha α and the generic transpose rank s s s ; the number, degrees and dimensions of the complex spans of the residue images; the relative multiplicities of the coefficient pencil and the signatures involved; and whether the transpose Gram vanishes on the singular stratum. They do not depend on the choice of orthonormal eigenframes at x x x : two choices differ by constant orthogonal block transformations commuting with Λ \Lambda Λ , which preserve every rank, dimension, support, degree and image inclusion.
Lemma 27.1 (Real-open branch). There is a finite partition U 0 = U 1 ⊔ ⋯ ⊔ U N U_0 = U_1 \sqcup\cdots\sqcup U_N U 0 = U 1 ⊔ ⋯ ⊔ U N into semialgebraic sets on each of which every invariant above is constant. At least one U j U_j U j has nonempty Euclidean interior.
Proof. All objects have bounded degree in ( x , y ) (x,y) ( x , y ) , so each invariant above takes only finitely many values. The geometric invariants (number, degrees, dimensions and spans of the residue images, α \alpha α and s s s ) are constructible functions of x x x [18 §9.7] , as in Remark 27.2 , and a constructible set defined over R \mathbb{R} R has a semialgebraic set of real points. Include the image and kernel frames at x x x as auxiliary real variables, constrained by the polynomial orthonormality and spectral equations; by frame independence, existential and universal quantification over the frames give the same predicate. Ranks and signatures are sign conditions on minors; a factorization of bounded degree is a system of coefficient equations with auxiliary real and imaginary coefficients; the absence of real points, the positive pencil member and its relative multiplicities are quantified polynomial conditions. So each value of each invariant is a first-order formula over R \mathbb{R} R in x x x , and by Tarski–Seidenberg elimination [4 Chapter 2] each such set is semialgebraic and the common refinement is finite. Outside the union of the zero sets of the finitely many nonzero polynomials in quantifier-free descriptions of the U j U_j U j , membership in each U j U_j U j is locally constant; a finite union of proper real zero sets contains no nonempty open set, so some point of U 0 U_0 U 0 has a neighbourhood inside one U j U_j U j . □ \square □
The branches need not be Zariski open, and signature boundaries may form lower-dimensional strata; the proof never uses them.
The cases of the partition
Every branch is covered by one of the results below. Each either classifies R R R directly, contradicts the branch, or gives full transpose alignment throughout the real-open region of the branch: the full transpose pencil has image E x = Im J x E_x=\operatorname{Im} J_x E x = Im J x for every real coefficient a ≠ 0 a\ne0 a = 0 , at every anchor x x x of the region.
(L1) ν = 7 \nu=7 ν = 7 : there is no pole at any anchor (Proposition 13.1).
(L2) ν = 8 \nu=8 ν = 8 , pole orders. The reduced denominator D D D satisfies D 2 ∣ det P x D^2\mid\det P_x D 2 ∣ det P x and has no real linear or odd-degree factor, so it is a quadratic Q Q Q , its square, a product Q 1 Q 2 Q_1Q_2 Q 1 Q 2 of distinct quadratics, or a real-irreducible quartic f f f (Proposition 11.3 ). Double quadratic poles are excluded (Proposition 15.1 ), as is a simple quadratic pole with residue rank four (Proposition 14.1 ). A quartic residue has rank two by Lemma 10.4 , and for Q 1 Q 2 Q_1Q_2 Q 1 Q 2 the same lemma, Q 1 r 1 Q 2 r 2 ∣ det P x Q_1^{r_1}Q_2^{r_2}\mid\det P_x Q 1 r 1 Q 2 r 2 ∣ det P x , gives r 1 = r 2 = 2 r_1=r_2=2 r 1 = r 2 = 2 . So every remaining pole is simple, with residue rank two.
(L3) Proper real support d ≤ 7 d\le7 d ≤ 7 . For a quadratic component, Q 2 Q^2 Q 2 divides the maximal d d d -minors, so d ≥ 4 d\ge4 d ≥ 4 ; d = 4 , 5 d=4,5 d = 4 , 5 are excluded componentwise (Lemma 17.1 ). For a sole quadratic, d = 6 d=6 d = 6 is excluded by Proposition 19.2 , and both d = 7 d=7 d = 7 alternatives, s = 4 s=4 s = 4 and s = 5 s=5 s = 5 , by Proposition 19.9 . For two quadratics, a component with d = 6 d=6 d = 6 is excluded by Proposition 19.2 , and d = 7 d=7 d = 7 by Proposition 19.5. A real-irreducible quartic pole always has d = 8 d=8 d = 8 (Lemma 17.3).
(L4) Full support. Incidence gives 4 ≤ s ≤ α ≤ 5 4\le s\le\alpha\le5 4 ≤ s ≤ α ≤ 5 ; the pair ( α , s ) = ( 5 , 4 ) (\alpha,s)=(5,4) ( α , s ) = ( 5 , 4 ) gives full transpose alignment (Proposition 17.5 ). In the remaining cases α = s \alpha=s α = s , and Lemma 17.6 makes distinct pole components have distinct residue images.
(L5) Two quadratic poles. Proposition 18.4 leaves two geometric quadratics or four geometric lines in one Galois orbit, so all geometric components have one type ( α , s ) (\alpha,s) ( α , s ) (Remark 27.2 ). Type ( 5 , 5 ) (5,5) ( 5 , 5 ) is excluded by the divisor budget (Proposition 21.2 ), and type ( 4 , 4 ) (4,4) ( 4 , 4 ) is aligned by Proposition 20.2 , whose hypotheses hold by Remark 20.9.
(L6) A single real pole of type ( 4 , 4 ) (4,4) ( 4 , 4 ) . By Remark 20.9, G G G vanishes on every image of type ( 4 , 4 ) (4,4) ( 4 , 4 ) , and full real support gives full complex span. A real geometrically integral image of full span has degree 4 4 4 , 6 6 6 or 8 8 8 , and each is treated in Proposition 20.3 . A conjugate pair of images aligns (Proposition 20.8 ).
(L7) A single real pole of type ( 5 , 5 ) (5,5) ( 5 , 5 ) with one real integral image V ( h , q ) V(h,q) V ( h , q ) , where q q q is positive and the maximal relative multiplicity of the pencil ( h , q ) (h,q) ( h , q ) is at most five (Propositions 21.3 and 23.1 ). The maximal multiplicity is at most three (Propositions 23.2 and 23.3 ), five (Propositions 22.4 and 24.6, using Lemma 24.2 ), or four (Propositions 22.4 and 24.6, using Lemmas 24.3 and 24.4 ).
The normal forms produced there align: unbalanced forms by Proposition 22.3, balanced forms by Proposition 22.2, which covers every relative multiplicity at most five; the extraction uses Lemma 12.8, Proposition 22.4 and Lemma 24.1.
(L8) A single real pole of type ( 5 , 5 ) (5,5) ( 5 , 5 ) with a conjugate pair of images V ( h , ℓ ) V(h,\ell) V ( h , ℓ ) , V ( h ˉ , ℓ ˉ ) V(\bar{h},\bar{\ell}) V ( h ˉ , ℓ ˉ ) (Proposition 21.3, Lemma 21.4 and Proposition 21.5). A transpose Gram vanishing on Σ = V ( h , ℓ ) ∩ V ( h ˉ , ℓ ˉ ) \Sigma= V(h,\ell) \cap V(\bar{h},\bar{\ell}) Σ = V ( h , ℓ ) ∩ V ( h ˉ , ℓ ˉ ) gives alignment (Proposition 25.1), and so does a constant column (Proposition 25.2, using Lemma 16.3). Otherwise Proposition 25.3 supplies complex linear columns, and Proposition 26.2 gives a global real balanced form, which aligns by Proposition 22.2.
The list is exhaustive. Items (L2)–(L4) reduce every pole at ν = 8 \nu= 8 ν = 8 to simple poles with rank-two residues, full real support, and type ( α , s ) = ( 4 , 4 ) (\alpha,s) = (4,4) ( α , s ) = ( 4 , 4 ) or ( 5 , 5 ) (5,5) ( 5 , 5 ) . The denominator has one real pole component (Q Q Q or f f f ) or two (Q 1 Q 2 Q_1Q_2 Q 1 Q 2 ), the latter being (L5). A single real pole component is geometrically irreducible or splits over C \mathbb{C} C into one conjugate pair, so it has one geometrically integral real image or a conjugate pair of images; the two conjugate components of a split real pole have the same ( α , s ) (\alpha,s) ( α , s ) , since B B B is real. With the two types this gives four alternatives: the two of (L6), and (L7) and (L8); in (L7) the maximal relative multiplicity takes one of the three listed values.
Proof of Theorem A
Proof. By the reductions above we may assume that Λ \Lambda Λ is indefinite. Suppose that R R R is not Clifford. Then F x F_x F x has a pole at every x ∈ U 0 x \in U_0 x ∈ U 0 , and ν = 8 \nu= 8 ν = 8 by (L1); the cases that conclude “Clifford” now give contradictions. By Lemma 27.1 some branch U j U_j U j contains a nonempty open set U ′ of anchors U' \text{ of anchors} U ′ of anchors . If the result of (L1)–(L8) covering U j U_j U j shows that R R R is Clifford, or that no anchor has the invariants of U j U_j U j , we have a contradiction. Otherwise it gives full transpose alignment: at every x ∈ U ′ x \in U' x ∈ U ′ the full transpose pencil has image E x E_x E x for every real a ≠ 0 a \ne0 a = 0 . By Lemma 16.2, this gives J y z = 0 \mathcal{J}_y z = 0 J y z = 0 for all y , z ∈ ker J x y,z \in\ker\mathcal{J}_x y , z ∈ ker J x , x ∈ U ′ x \in U' x ∈ U ′ . Lemma 16.1 applies on the open set U ′ U' U ′ and makes the full Gram matrix of the coherent partner radial, hence polynomial. Since Λ \Lambda Λ is nonscalar, Proposition 9.1(2) shows that R R R is Clifford, a contradiction.
Acknowledgments
This work was done overwhelmingly by AI models and supervised by the author. The models were Anthropic’s Claude and OpenAI’s ChatGPT. The author takes full responsibility for the paper.
Statements and Declarations
Funding. No funding was received for conducting this study.
Competing interests. The author has no competing interests to declare that are relevant to the content of this article.
Data availability. No datasets were generated or analysed during the current study. The Lean files are available at the GitHub link in the footnote on the first page; the expanded version is [17 ] .