\section{An extreme Jacobi eigenvalue of multiplicity seven or eight}\label{sec:extreme} \begin{mainthm}\label{thm:extreme} Let $R$ be an Osserman algebraic curvature tensor on $\R^{16}$ 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 $R$ is Clifford. \end{mainthm} \noindent 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 $R_1$, so that the reduced Jacobi operator has nonzero rank $\nu\le8$, and $\Lambda$ is the diagonal matrix of the nonzero Jacobi eigenvalues for unit vectors, with multiplicities. \begin{proposition}\label{prop:gramconditional} Let $R$ be an Osserman algebraic curvature tensor on $\R^{16}$ as above. \begin{enumerate} \item If $\Jac_x=M_x\Lambda M_x^T$ has a global linear factor $M_x:\R^\nu\to\R^{16}$, injective for all $x\ne0$, then $R$ is Clifford. \item Suppose that $\Jac_z=M_z\Lambda M_z^T$ has a rational homogeneous factor of degree one with the following coherence property: there are a real unit vector $X$ and an orthonormal eigenframe $M_X$ such that, for all $Y\perp X$, \begin{equation}\label{eq:shear} M_{sX+Y}=sM_X+M_Y \end{equation} as a rational identity. If the \emph{full} Gram matrix $\Phi(z)=M_z^TM_z$ on $\R^{16}$ is polynomial, then $R$ is Jacobi-proportional, hence, by Corollary~\ref{cor:C}, Clifford or a linear combination of $R_1$ and an orthogonal conjugate of $R_{\rm Cay}$. If $\Lambda$ is nonscalar, it is Clifford. \end{enumerate} \end{proposition} \noindent In part~(2) the full Gram is $\Phi(sX+Y)=s^2I+s(M_X^TM_Y+M_Y^TM_X)+M_Y^TM_Y$, so its mixed coefficient must be polynomial as well. \subsection{Quadratic Gram rigidity} \begin{lemma}\label{lem:gramrigidity} Let $\Lambda$ be invertible symmetric of size $\nu\le8$. Suppose $\Phi(x)$ is a symmetric $\nu\times\nu$ matrix, homogeneous quadratic in $x\in\R^{16}$, positive definite for $x\ne0$, and $\Lambda\Phi(x)$ has the spectrum of $\norm x^2\Lambda$, with multiplicities. If $\Phi(x_0)=I$ at a unit vector $x_0$, then $\Phi(x)=\norm x^2I$ for every $x$. \end{lemma} \begin{proof} This refines \cite[Section~5, Lemmas~5--8, pp.~11--15 of arXiv:math/0204258v2]{Nik2005}, whose hypothesis $n>(\nu+1)^2/4$ fails for $n=16$, $\nu=7,8$; the remaining pairs are settled by a topological count. Diagonalize $\Lambda$ orthogonally, choose an eigenvalue $\lambda$ of multiplicity $m$ whose reciprocal is extremal, and set $q=\norm x^2$, $Q_\lambda(x)=\lambda q\Lambda^{-1}-\Phi(x)$. For unit $x$ the matrix $\Phi(x)^{1/2}\Lambda\Phi(x)^{1/2}$ is symmetric with the spectrum of $\Lambda$; write it as $O\Lambda O^T$ with $O$ orthogonal. Then $C=O^T\Phi(x)^{1/2}$ satisfies $C\Lambda C^T=\Lambda$, $\Phi=C^TC$ and $Q_\lambda(x)=C^T(\lambda\Lambda^{-1}-I)C$, so $Q_\lambda(x)$ has kernel dimension $m$ for all $x$ and is semidefinite of a fixed sign. For $u\in\R^\nu$ let $H_\lambda(u)$ be the symmetric $16\times16$ matrix with $x^TH_\lambda(u)x=u^TQ_\lambda(x)u$; it is quadratic in $u$ and semidefinite, so $L_u:=\{x:Q_\lambda(x)u=0\}=\ker H_\lambda(u)$ is a linear subspace, of dimension $d(u)$. Choose independent $u_1,\ldots,u_m$ maximizing $\sum_i d(u_i)$ and put $S=\Span(u_1,\ldots,u_m)$. For $u\notin S$ and each $i$, the vectors $u$ and $u_j$ ($j\ne i$) are independent, so $d(u)\le d(u_i)$. For any $m+1$ independent coefficient vectors $v_0,\ldots,v_m$ one has $L_{v_0}\cap\cdots\cap L_{v_m}=0$ (otherwise $\ker Q_\lambda(x)$ would have dimension at least $m+1$ at some $x\ne0$), and the subspace intersection inequality $0=\dim(L_u\cap L_{u_1}\cap\cdots\cap L_{u_m})\ge d(u)+\sum_id(u_i)-16m$, with $d(u)\le d(u_i)$ for each $i$, gives $(m+1)d(u)\le d(u)+\sum_id(u_i)\le16m$, that is, \begin{equation}\label{eq:unrounded} d(u)\le h:=\left\lfloor\frac{16m}{m+1}\right\rfloor \qquad(u\notin S). \end{equation} The incidence space $\mathcal I=\{(x,u):x\ne0,\ Q_\lambda(x)u=0\}$ is a connected real-analytic rank-$m$ vector bundle over $\R^{16}\setminus\{0\}$, of dimension $16+m$, and its open subset $\mathcal I_{\rm out}$ where $u\notin S$ has semialgebraic dimension at most $\nu+h$. If $\nu+h<16+m$, then $\mathcal I_{\rm out}$ is empty and $\ker Q_\lambda(x)=S$ for every $x$. The scalar case $m=\nu$ is immediate, and for $\nu\le8$ the only remaining nonscalar pairs are \[ \begin{array}{c|rr} (\nu,m)&h&\nu+h\\ \hline (7,3)&12&19\\ (8,2)&10&18\\ (8,3)&12&20\\ (8,4)&12&20\\ (8,5)&13&21 \end{array} \] Suppose $\mathcal I_{\rm out}\ne\varnothing$ in one of these cases and put $\ell=16+m-\nu$. The coefficient strata with $d(u)\le\ell-1$ have incidence dimension at most $\nu+\ell-1=15+m$, so the locus with $d(u)\ge\ell$ contains a relatively open subset of $\mathcal I$. All $(17-\ell)$-minors of $H_\lambda(u)$ vanish there; their pullbacks to $\mathcal I$ are real analytic, so by connectedness they vanish on $\mathcal I$, and every nonempty coefficient fibre has $d(u)\ge\ell$. Normalize both variables: \[ \mathcal E=\{(x,u)\in S^{15}\times S^{\nu-1}:Q_\lambda(x)u=0\}, \qquad Z=\{u\in S^{\nu-1}:\det H_\lambda(u)=0\}. \] The first projection makes $\mathcal E$ an $S^{m-1}$-bundle over $S^{15}$, so $\dim\mathcal E=14+m$. The second projection is a proper surjection onto the compact real algebraic set $Z$, with fibres $S(L_u)$ of dimension at least $\ell-1$, and at most $h-1$ outside $S$ by \eqref{eq:unrounded}. As $\mathcal E_{\rm out}$ is nonempty and open, \[ 15+m-h\ \le\ z:=\dim Z\ \le\ 15+m-\ell=\nu-1 , \] which gives \[ \begin{array}{c|rcc} (\nu,m)&\ell&\text{possible }z&\text{minimum sphere-fibre dimension}\\ \hline (7,3)&12&6&11\\ (8,2)&10&7&9\\ (8,3)&11&6\text{ or }7&10\\ (8,4)&12&7&11\\ (8,5)&13&7&12 \end{array} \] In every row $z<\ell-1$, so the finite-degree Vietoris--Begle theorem \cite[Theorem~1.1 (Begle), p.~383]{Kryszewski}, with mod-two coefficients, gives $H^z(Z;\mathbb F_2)\cong H^z(\mathcal E;\mathbb F_2)$ (all spaces are compact semialgebraic and triangulable, so \v Cech and singular cohomology agree). The latter group vanishes by the Gysin sequence of the $S^{m-1}$-bundle over $S^{15}$: the base groups in degrees $z$ and $z-m+1$ vanish in every row. But a nonempty compact real algebraic set of dimension $z$ has a nonzero top mod-two homology class \cite[Theorem~1.1, p.~22]{BochnakKucharz}, hence $H^z(Z;\mathbb F_2)\ne0$, a contradiction. Hence $\ker Q_\lambda(x)=S$ for all $x\ne0$; at $x_0$ this is the $\lambda$-eigenspace of $\Lambda$, so $\Phi(x)u=\lambda q\Lambda^{-1}u=qu$ for $u\in S$. Symmetry removes the cross blocks, and restricting $\Phi,\Lambda$ to the orthogonal complement preserves all hypotheses. Induction on $\nu$ proves $\Phi=qI$. \end{proof} \subsection{Proof of Proposition~\ref{prop:gramconditional}} \begin{proof}[Proof of part~(1)] A constant coefficient change $M_x\mapsto M_xG_0$ normalizes $M_{x_0}^TM_{x_0}=I$ at one unit vector and replaces $\Lambda$ by $G_0^{-1}\Lambda G_0^{-T}$; this replaces $\Lambda\Phi(x)$ by the similar matrix $G_0^{-1}\Lambda\Phi(x)G_0$, and at $x_0$ the new $\Lambda$ has the spectrum of the old one, so the hypotheses of Lemma~\ref{lem:gramrigidity} are kept. For a global linear factor, $\Phi(x)=M_x^TM_x$ is homogeneous quadratic and positive definite, and the nonzero spectrum of $\Jac_x$ is that of $\Lambda\Phi(x)$, so Lemma~\ref{lem:gramrigidity} gives $M_x^TM_x=qI$. For a coefficient basis vector $e_i$ write $J_ix=M_xe_i$. Since $\Jac_xx=0$, $M_x$ is injective and $\Lambda$ is invertible, $M_x^Tx=0$, so each $J_i$ is skew; the Gram identity makes the $J_i$ orthogonal with $J_iJ_j+J_jJ_i=-2\delta_{ij}I$ by polarization. Diagonalizing $\Lambda$ gives the Clifford formula (with no columns, $R$ has constant curvature). \end{proof} \begin{proof}[Proof of part~(2)] On the real regular locus $\Phi=M^TM$ is positive definite, and the nonzero spectrum gives $\det\Phi(z)=q(z)^\nu$. If $\Phi$ is polynomial, it is positive semidefinite everywhere by continuity and positive definite for $z\ne0$ by the determinant. By \eqref{eq:shear} and homogeneity, $M_{X+tY}=M_X+tM_Y$ for generic $Y\perp X$ and $t\ne0$, and continuity at $t=0$ gives $\Phi(X)=I$. Lemma~\ref{lem:gramrigidity} yields $\Phi(z)=q(z)I$. For a nonzero eigenvalue $\lambda_j$ let $I_j$ be its coefficient projector and $\pi_j(z)$ the corresponding reduced Jacobi projector. Then $C_j(z):=M_zI_jM_z^T=q(z)\pi_j(z)$ on the regular locus, hence as a rational identity. In coordinates $z=sX+Y$ the left side has a denominator dividing a power of a polynomial $D(Y)$ independent of $s$, by \eqref{eq:shear}, while the right side has only powers of $q(z)=s^2+\norm Y^2$ as possible denominators: by Lagrange interpolation over the distinct nonzero eigenvalues $\lambda_1,\ldots,\lambda_p$, $\pi_j(z)=\Jac_z\prod_{k\ne j}(\Jac_z-\lambda_kq(z)I)\big/ \bigl(\lambda_j\prod_{k\ne j}(\lambda_j-\lambda_k)\,q(z)^p\bigr)$. The polynomial $q(z)$ is irreducible over $\mathbb C$ in sixteen variables and coprime to every nonzero $D(Y)$, so unique factorization shows that $C_j$ is polynomial, and homogeneity makes it quadratic; so is the reduced zero-eigenvalue family $C_0(z)=q(z)I-zz^T-\sum_jC_j(z)$. Each $C_j$ is symmetric with $C_j(z)z=0$, and the polarization argument of Section~\ref{sec:quartic} gives $\ip{C_j(x)y}y=\ip{C_j(y)x}x$. So $R$ is Jacobi-proportional, and Corollary~\ref{cor:C} applies. If $\Lambda$ is nonscalar, the zero eigenvalue and at least two distinct nonzero eigenvalues give three reduced eigenvalues, which excludes the Cayley alternative. \end{proof} \subsection{The coherent partner} \begin{lemma}\label{lem:rationalpartner} For the shifted tensor above, with $\nu\le8$, there are a real unit anchor $X$, an orthonormal eigenframe $M_X$, and a rational homogeneous partner $M_Y$ on $X^\perp$ with \[ \Jac_{sX+Y}=(sM_X+M_Y)\Lambda(sM_X+M_Y)^T. \] \end{lemma} \begin{proof} We use two results of \cite[Section~4]{Nik2005}, both stated for $n\ge2\nu$, hence valid for $n=16$, $\nu\le8$. By \cite[Lemma~2(1) and its proof, p.~7 of arXiv:math/0204258v2]{Nik2005}, for every unit $X$ the unit vectors $Y\perp X$ with $\operatorname{Im}\Jac_X\cap\operatorname{Im}\Jac_Y=0$ form an open dense subset of the unit sphere of $X^\perp$. By \cite[Lemma~3, p.~8 of arXiv:math/0204258v2]{Nik2005}, for every such orthonormal pair $(X,Y)$ and every $M_X$ with $\Jac_X=M_X\Lambda M_X^T$ there is a unique $M_Y$ with $\Jac_{sX+tY}=(sM_X+tM_Y)\Lambda(sM_X+tM_Y)^T$ for all real $s,t$. Fix a unit $X$ and an orthonormal eigenframe $M_X$, and put $M_{cY}=cM_Y$ for $c\ne0$; uniqueness makes this consistent. Write $M_Y=M_XF(Y)+K_XP(Y)$, where $K_X$ is an orthonormal frame of $\ker\Jac_X$, and $2\mathcal B(X,Y)=\Jac_{X+Y}-\Jac_X-\Jac_Y$. The mixed coefficient gives $P(Y)=K_X^T\,2\mathcal B(X,Y)M_X\Lambda^{-1}$, which is linear, of full column rank on the generic-pair locus since the images of $\Jac_X$ and $\Jac_Y$ meet only in $0$. The lower-upper block of $\Jac_Y$ gives $P\Lambda F^T=K_X^T\Jac_YM_X$, and the rational left inverse $(P^TP)^{-1}P^T$ shows that $F$, hence $M_Y$, is rational homogeneous of degree one. The two-plane identities hold rationally by continuation, and defining $M_{sX+Y}=sM_X+M_Y$ gives the required coherent factor. \end{proof} \subsection{Proof of Theorem~\ref{thm:extreme}} \begin{proof} Shift the extreme eigenvalue of multiplicity seven or eight to zero; then $\nu=7$ or $8$ and $\Lambda$ is definite. Use Lemma~\ref{lem:rationalpartner} with coordinates $Y=(y_1,\ldots,y_{15})$ on $X^\perp$, and put $\mathbb F=\R(y_1,\ldots,y_{15})$, $d=\sum_iy_i^2$, $G(s)=s^2I+s\Phi_1+\Phi_0$, where $\Phi_1=M_X^TM_Y+M_Y^TM_X$ and $\Phi_0=M_Y^TM_Y$ have entries in $\mathbb F$; thus $G(s)=\Phi(sX+Y)$ is the full Gram of Proposition~\ref{prop:gramconditional}. Equality of the nonzero Jacobi spectra gives \begin{equation}\label{eq:spectralpencil} \det(tI-\Lambda G(s))=\prod_{i=1}^{\nu}\bigl(t-(s^2+d)\mu_i\bigr) \quad\text{in }\mathbb F[s,t], \end{equation} where the $\mu_i$ are the nonzero eigenvalues with multiplicities. Let $\mathbb E=\mathbb F(\alpha)$ with $\alpha^2=-d$; as $\mathbb F$ is formally real, $[\mathbb E:\mathbb F]=2$, and $\mathbb E$ is the fraction field of $\R[s,y_1,\ldots,y_{15}]/(s^2+y_1^2+\cdots+y_{15}^2)$. By \cite[Lemma~3.11 and its proof, p.~398]{Nik2006}, a sum of at most eight squares in $\mathbb E$ vanishes only when every term vanishes, so every constant definite real quadratic form of dimension at most eight is anisotropic over $\mathbb E$. At $s=\alpha$, \eqref{eq:spectralpencil} shows that $N=\Lambda G(\alpha)$ is nilpotent, and it is self-adjoint for the anisotropic form $\langle v,w\rangle_{\Lambda^{-1}}=v^T\Lambda^{-1}w$, since $\Lambda^{-1}N=G(\alpha)$ is symmetric. Such an $N$ is zero: if its nilpotency index were $e\ge2$, then $w=N^{e-1}v\ne0$ for some $v$ and $\langle w,w\rangle_{\Lambda^{-1}}=\langle v,N^{2e-2}v\rangle_{\Lambda^{-1}}=0$, contradicting anisotropy. As $\Lambda$ is invertible, $\Phi_0-dI+\alpha\Phi_1=0$, and the independence of $1,\alpha$ over $\mathbb F$ gives $\Phi_1=0$ and $\Phi_0=dI$. So the full Gram is $G(s)=(s^2+d)I$, polynomial on $\R^{16}$, and Proposition~\ref{prop:gramconditional}(2) makes $R$ Clifford (there are at least three reduced eigenvalues). \end{proof}