\section{Further multiplicity reductions}\label{sec:smallroots} Theorems~\ref{thm:A} and~\ref{thm:D} complete several polynomial-eigenfield reductions of \cite{Nik2006}. \begin{mainthm}\label{thm:E} An Osserman algebraic curvature tensor on $\R^{16}$ is Clifford if its reduced multiplicity pattern is one of \[ (7,6,1,1),\ (8,5,1,1),\ (7,5,2,1),\ (8,4,2,1),\ (7,5,1,1,1),\ (8,4,1,1,1), \] or if one eigenvalue has multiplicity seven or eight and every other eigenvalue has multiplicity at most two (nine further patterns, fifteen in all). \end{mainthm} \subsection{Constant projections onto Cayley eigenspaces} \begin{lemma}\label{lem:constantprojection} Let $E(x),N(x)$ be the two Cayley distributions of Lemma~\ref{lem:nofield}. If a skew-symmetric linear map $K$ satisfies $\norm{\pi_{N(x)}Kx}^2=c\norm x^2$ for all $x\ne0$ and a constant $c$, then $K=0$. In particular an orthogonal skew map cannot have constant normalized projection lengths onto the two Cayley eigenspaces. \end{lemma} \begin{proof} Use the symmetric Clifford system of Section~\ref{sec:Cayley}, written here $H_0,\ldots,H_8$ (there $S_0,\ldots,S_8$), and put $q=\norm x^2$, $h_i(x)=\ip{H_ix}x$, $a_i(x)=\ip{H_ix}{Kx}$. The Clifford relations give $\ip{H_ix}{H_jx}=\delta_{ij}q$. The identity $\sum_i(H_ix)(H_ix)^T=Z_x=xx^T+qI-Q(x)$ of Section~\ref{sec:Cayley} (with $H_i=S_i$) and $Q(x)=q\,\pi_{L(x)}$ (Section~\ref{sec:quartic}), with $x\in L(x)$, give $\sum_i(H_ix)(H_ix)^T=xx^T+q\,\pi_{N(x)}=q\,\pi_{\R x+N(x)}$. So the nine orthogonal vectors $H_ix$, of norm $\sqrt q$, span $\R x+N(x)$, and, $K$ being skew, $\sum_i a_i^2=cq^2$ and $\sum_i h_i^2=q^2$; differentiating the latter along $Kx$ gives $\sum_i h_ia_i=0$. The matrix of $a_i$ is the symmetric trace-free $A_i=(H_iK-KH_i)/2$. If $c>0$, put $H_i(t)=\cos t\,H_i+c^{-1/2}\sin t\,A_i$. Then $\sum_i\ip{H_i(t)x}x^2=q^2$ for every $t$, and this quadratic map is not constant on the unit sphere (otherwise $H_i(t)=b_iI$ with $b$ a unit vector, but all its matrices are trace-free). By Lemma~\ref{lem:threshold}(c) the $H_i(t)$ form a symmetric Clifford system, so $\Omega(t)=H_0(t)\cdots H_8(t)$ is a symmetric involution, and $\tr\Omega(t)$ is continuous and integer-valued, hence constant. But $\Omega(0)=\pm I$ on $\R^{16}$ (as nine is odd, $\Omega(0)$ commutes with every $H_i$, and, by Section~\ref{sec:Cayley}, the products of the $H_i$ span the symmetric matrices, so a matrix commuting with every $H_i$ is scalar) and $\Omega(\pi)=-\Omega(0)$, a contradiction. If $c=0$, then $Kx\in L(x)$, and skewness gives $Kx\in E(x)$, so $K=0$ by Lemma~\ref{lem:nofield}. \end{proof} \subsection{Polynomial tools and a three-column reduction} From here to the end of this section, $A(x)$ denotes the Jacobi operator $\Jac_x$ after one reduced eigenvalue has been shifted to zero (not the normalized family of Section~\ref{sec:quartic}), $q=\norm x^2$, $\mathcal O=\R[x_1,\ldots,x_{16}]/(q)$, and $\pi\colon\R[x_1,\ldots,x_{16}]\to\mathcal O$ is the projection. As $q$ is irreducible, it is a prime element of $\R[x_1,\ldots,x_{16}]$. We use two results of \cite{Nik2006}: $\mathcal O$ is a UFD \cite[Theorem~3.1, p.~397]{Nik2006}; and if $q$ divides a sum of at most eight real polynomial squares, it divides every polynomial being squared \cite[Lemma~3.11, p.~398]{Nik2006}. A polynomial eigenvector field of a simple eigenvalue is a \emph{simple eigenfield}. The simple eigenfields of \cite[Lemma~3.1(1), p.~383]{Nik2006} satisfy \begin{equation}\label{eq:cubicidentities} \ip{P(x)}x=0,\quad \norm{P(x)}^2=q^3,\quad P(P(x))=-q^4x, \end{equation} after multiplication by $q$ if they are linear: equation~(7) of that lemma states them for unit $x$, and homogeneity gives them for all $x$. We also use \cite[Lemma~3.5, pp.~388--390]{Nik2006}. It is stated there under the standing hypothesis of \cite[Subsection~3.1]{Nik2006}, a Jacobi eigenvalue of multiplicity nine, and for an eigenvector field $P$; its proof uses neither. We state it in the form used here, with that proof. \begin{lemma}\label{lem:cubic-gcd} Let $P\colon\R^{16}\to\R^{16}$ have homogeneous cubic components that satisfy \eqref{eq:cubicidentities}. If the images in $\mathcal O$ of the sixteen components of $P$ have a nonunit common divisor, then $P(x)=qJx$ for an orthogonal skew map $J$. \end{lemma} \begin{proof} We follow \cite[pp.~389--390]{Nik2006}. If every component of $P$ is divisible by $q$, then $P(x)=qJx$ with $J$ linear, and \eqref{eq:cubicidentities} gives $\ip{Jx}x=0$ and $\norm{Jx}^2=q$, so $J$ is skew and orthogonal. Otherwise the common divisor is homogeneous, as a divisor of a nonzero homogeneous element of the graded domain $\mathcal O$; its degree $r$ is positive, since the units of $\mathcal O$ are the nonzero constants, and $r\le3$, since it divides a nonzero cubic. Lifting to $\R[x_1,\ldots,x_{16}]$, we get $P(x)=\varphi(x)\Psi(x)+qCx$ with $\varphi$ homogeneous of degree $r$ and not divisible by $q$, the components of $\Psi$ homogeneous of degree $3-r$, and $C$ linear. It suffices to show that every component of $\Psi$ is divisible by $q$. If $r=3$, then $\Psi$ is constant and $\varphi\ip \Psi x=-q\ip{Cx}x$. As $q$ is prime and does not divide $\varphi$, it divides the linear form $\ip \Psi x$, so $\Psi=0$. If $r=2$, then $\Psi(x)=Bx$ with $B$ linear. Modulo $q$, $\norm{P(x)}^2\equiv \varphi(x)^2\norm{Bx}^2$, so $q$ divides $\norm{Bx}^2$, and $B=cU$ with $U$ orthogonal and $c\ge0$. Since $\norm{P(x)}^2=q^3$ and $P(x)\equiv \varphi(x)Bx$ modulo $q$, we get \[ 0\equiv P(P(x))\equiv \varphi(P(x))\,BP(x)\equiv \varphi(x)^3\varphi(Bx)\,B^2x \pmod q. \] If $c\ne0$, then $B^2$ is invertible, so $q$ divides $\varphi(Bx)$, and the substitution $x=B^{-1}y$ shows that $c^{-2}\norm y^2$ divides $\varphi(y)$, which is false. So $c=0$ and $\Psi=0$. If $r=1$, then $\varphi$ is a nonzero linear form, and $\varphi\ip \Psi x=-q\ip{Cx}x$ shows that $\varphi$ divides $\ip{Cx}x$, say $\ip{Cx}x=\varphi(x)\ip ax$ with $a\in\R^{16}$. Replacing $C$ by $\tilde Cx=Cx-\varphi(x)a$ and $\Psi$ by $\tilde \Psi=\Psi+qa$ keeps $P=\varphi\tilde \Psi+q\tilde Cx$ and makes $\tilde C$ skew. Modulo $\varphi$, the identity $\norm{P(x)}^2=q^3$ shows that $\varphi$ divides $q-\norm{\tilde Cx}^2=x^T(I+\tilde C^2)x$. 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 $I+\tilde C^2$ has even multiplicity, $I+\tilde C^2=0$, and $\tilde C$ is orthogonal. The third identity in \eqref{eq:cubicidentities} now reads $\varphi(P(x))\tilde \Psi(P(x))=-q^3\varphi(x)\tilde C\tilde \Psi(x)$. On the hyperplane $\varphi=0$ we have $P(x)=q\tilde Cx$, so $\varphi(\tilde Cx)\tilde \Psi(\tilde Cx)$ vanishes there. Writing $\varphi(x)=\ip fx$, the form $\varphi(\tilde Cx)=-\ip{\tilde Cf}x$ does not vanish identically on $f^\perp$, since $\tilde Cf\ne0$ is orthogonal to $f$. Hence $\varphi$ divides every component of $\tilde \Psi(\tilde Cx)$, and, as $\tilde C^{-1}=-\tilde C$, we get $\tilde \Psi(x)=\varphi(\tilde Cx)B'x$ with $B'$ linear. Then $P(x)=\varphi(x)\varphi(\tilde Cx)\,B'x+q\tilde Cx$, where $\varphi(x)\varphi(\tilde Cx)$ is a quadratic form not divisible by $q$, and the case $r=2$ applies: $B'=0$, so $\tilde \Psi=0$, and $\Psi=-qa$ is divisible by $q$. \end{proof} The three-column step of \cite[pp.~387--389]{Nik2006} reads as follows. Let $P_1,P_2,P_3$ be mutually orthogonal cubic eigenfields of squared norm $q^3$ satisfying $\sum_{i=1}^3w_iP_iP_i^T\equiv0$ modulo $q$, with $w_1,w_2>0>w_3$ (identities of this kind come from the spectral projectors, as in \eqref{eq:twocolumn} below). After rescaling the columns by nonzero constants, their images $p_i\in\mathcal O^{16}$ satisfy \begin{equation}\label{eq:rankoneK} p_1p_1^T+p_2p_2^T=p_3p_3^T. \end{equation} The right side of \eqref{eq:rankoneK} has rank at most one, so every $2\times2$ minor of the left side vanishes; by Lagrange's identity its principal minor in rows and columns $a,b$ is $(p_{1a}p_{2b}-p_{2a}p_{1b})^2$, so, $\mathcal O$ being a domain, all the minors $p_{1a}p_{2b}-p_{2a}p_{1b}$ of $[p_1\ p_2]$ vanish. If some $p_i=0$, then $P_i$ is $q$ times a linear eigenfield. Otherwise unique factorization gives a primitive $v$ with $p_1=u_1v$, $p_2=u_2v$; then \eqref{eq:rankoneK} gives $p_3=u_3v$ and $u_1^2+u_2^2=u_3^2$: a nonzero component of $v$ gives a square root of $u_1^2+u_2^2$ in the fraction field, integral over the integrally closed ring $\mathcal O$, hence in $\mathcal O$; cross entries fix one sign in all components. Suppose the negative column comes from a simple eigenvalue. If $u_3$ is a nonunit, Lemma~\ref{lem:cubic-gcd} makes $P_3/q$ linear. If $u_3$ is a unit, homogeneity makes $p_1,p_2$ constant multiples of $p_3$, and a constant combination of the original fields satisfies \begin{equation}\label{eq:linearcombination} c_1P_1(x)+c_2P_2(x)=qJx,\qquad c_1^2+c_2^2=1, \end{equation} with $J$ orthogonal and skew (by norm and tangency). If $P_1,P_2$ have the same eigenvalue, this is a linear eigenfield. Otherwise $Jx$ has constant coordinates in their moving orthonormal frame on the unit sphere, so subtracting $CR_J$, for any constant $C$, leaves an Osserman tensor. \subsection{Two simple eigenvalues}\label{subsec:twosimple} Consider $(M,N,1,1)$ with $(M,N)=(7,6)$ or $(8,5)$. Shift the eigenvalue of multiplicity $M$ to zero; let $\lambda$ be the eigenvalue of multiplicity $N$ and $\mu,\sigma$ the simple eigenvalues (distinct and nonzero). By \cite[Lemma~3.1]{Nik2006} there are cubic normalized eigenfields $P_1,P_2$ (linear ones homogenized by $q$), and the spectral projectors give \begin{equation}\label{eq:twocolumn} \mu(\mu-\lambda)P_1P_1^T+\sigma(\sigma-\lambda)P_2P_2^T=qA(A-\lambda qI). \end{equation} If the two coefficients have the same sign, the sum-of-squares lemma on the diagonal makes both fields $q$ times linear ones. Moving one simple eigenvalue into each large eigenvalue gives the pattern $(M+1,N+1)$, that is $(8,7)$ or $(9,6)$: the first is classified by Theorem~\ref{thm:A}, with Cayley excluded by Lemma~\ref{lem:nofield}; the second is Clifford by \cite[Proposition~2.1, p.~382]{Nik2006}. Restore the two rank-one terms with \cite[Lemma~3.2(1)]{Nik2006}, first the one moved into $\lambda$ and then the one moved into $0$. A Clifford structure with $\nu=8$ (the $J_i$ of $\eta_i=0$ dropped) gives its base eigenvalue multiplicity exactly $15-\nu=7$, the dimension of $x^\perp\cap(\sum_i\R J_ix)^\perp$. The intermediate pattern $(M+1,N,1)$, that is $(8,6,1)$ or $(9,5,1)$, has no eigenvalue of multiplicity seven, so its Clifford structures have $\nu\le7$, and Remark~\ref{rem:lemma32-cl8} is needed only at an endpoint of type \eqref{eq:Cl8}. This proves the claim. Otherwise, after interchanging $\mu,\sigma$, put $a^2=\mu(\mu-\lambda)>0$ and $b^2=-\sigma(\sigma-\lambda)>0$, and let $p_i$ be the image of the unscaled $P_i$ in $\mathcal O^{16}$. Modulo $q$, \eqref{eq:twocolumn} gives $(ap_1)(ap_1)^T=(bp_2)(bp_2)^T$, so $ap_1=\pm bp_2$ (compare a nonzero component, then every cross entry). If both vanish, the linear-field argument above applies. Otherwise, after changing the sign of $P_2$, \[ aP_1-bP_2=\sqrt{a^2+b^2}\,qJx \] with $J$ orthogonal and skew. Put $C=\mu+\sigma-\lambda$; it is nonzero since $a^2+b^2=(\mu-\sigma)C$. On the unit sphere, subtracting $C(Jx)(Jx)^T$ turns the two-dimensional block into a fixed symmetric matrix of trace $\lambda$ and determinant $\mu\sigma-C(\sigma a^2+\mu b^2)/(a^2+b^2)=0$, since $\sigma a^2+\mu b^2=\mu\sigma(\mu-\sigma)$. The new eigenvalues are $0,\lambda$, so the pattern is $(M+1,N+1)$, and the projection lengths of $Jx$ onto the new eigenspaces are constant, because the changed block is a constant matrix in the frame $(P_1,P_2)$. Lemma~\ref{lem:constantprojection} excludes a Cayley endpoint (for $(M,N)=(8,5)$ the endpoint $(9,6)$ is Clifford by \cite[Proposition~2.1]{Nik2006}). So the endpoint is Clifford, and \cite[Lemma~3.2(1)]{Nik2006}, with Remark~\ref{rem:lemma32-cl8} at an endpoint of type \eqref{eq:Cl8}, restores the original tensor. \subsection{A double and a simple eigenvalue} Consider $(M,N,2,1)$ with $(M,N)=(7,5)$ or $(8,4)$. Shift $M$ to zero and call the other eigenvalues $\lambda,\mu,\sigma$, with $\mu$ double and $\sigma$ simple. There are cubic normalized fields $P_1,P_2$ for $\mu$ and $P_3$ for $\sigma$ with \[ \mu(\mu-\lambda)(P_1P_1^T+P_2P_2^T)+\sigma(\sigma-\lambda)P_3P_3^T =qA(A-\lambda qI). \] For equal coefficient signs all fields are linear after division by $q$. For opposite signs, after multiplying by $-1$ 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 \eqref{eq:linearcombination}, and a column vanishing modulo $q$ is already a linear field. Moving that field into the zero eigenvalue gives the endpoints \[ \begin{array}{c|cc} \text{initial}&\text{field from simple eigenvalue}&\text{field from double eigenvalue}\\ \hline (7,5,2,1)&(8,5,2)&(8,5,1,1)\\ (8,4,2,1)&(9,4,2)&(9,4,1,1) \end{array} \] which are Clifford by Theorem~\ref{thm:D}, Subsection~\ref{subsec:twosimple} or \cite[Proposition~2.1, p.~382]{Nik2006}. Restore the rank-one term by \cite[Lemma~3.2(1)]{Nik2006}; no endpoint in the table has an eigenvalue of multiplicity seven, so their Clifford structures have $\nu\le7$ and Remark~\ref{rem:lemma32-cl8} is not needed. \subsection{Three simple eigenvalues} Consider $(M,N,1,1,1)$ with $(M,N)=(7,5)$ or $(8,4)$. After shifting $M$ to zero there are five reduced eigenvalues, so \cite[Lemma~3.1]{Nik2006} gives fields of odd degree at most three. Let $\lambda$ be the eigenvalue of multiplicity $N$ and $\mu_1,\mu_2,\mu_3$ the simple eigenvalues; the cubic spectral identity has weights $w_i=\mu_i(\mu_i-\lambda)$. If all signs agree, every field is linear after division by $q$, 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 $(8,5,1,1)$ or $(9,4,1,1)$, Clifford by Subsection~\ref{subsec:twosimple} and \cite[Proposition~2.1]{Nik2006}; restore by \cite[Lemma~3.2(1)]{Nik2006}. Otherwise \eqref{eq:linearcombination} mixes $\mu_1,\mu_2$. Put $d=\mu_1c_2^2+\mu_2c_1^2$. After subtracting $C(Jx)(Jx)^T$ the characteristic equation of the block is $(\mu_1-z)(\mu_2-z)-C(d-z)=0$. Choose $z\in\{0,\lambda\}$ with $d\ne z$ (possible since $\lambda\ne0$) and $C=(\mu_1-z)(\mu_2-z)/(d-z)$, which is nonzero. The other eigenvalue is $\theta=\mu_1+\mu_2-C-z$, and $0,\lambda,\mu_3$ survive unchanged and distinct, so there is no two-eigenvalue endpoint. The complete list of endpoints, including coincidences, is as follows, where \emph{other large} means $\theta=\lambda$ if $z=0$ and $\theta=0$ if $z=\lambda$: \begin{center} \small \begin{tabular}{cc|cccc} $(M,N)$&$z$&$\theta$ new&$\theta=\mu_3$&$\theta$ other large&$\theta=z$\\ \hline $(7,5)$&$0$&$(8,5,1,1)$&$(8,5,2)$&$(8,6,1)$&$(9,5,1)$\\ $(7,5)$&$\lambda$&$(7,6,1,1)$&$(7,6,2)$&$(8,6,1)$&$(7,7,1)$\\ $(8,4)$&$0$&$(9,4,1,1)$&$(9,4,2)$&$(9,5,1)$&$(10,4,1)$\\ $(8,4)$&$\lambda$&$(8,5,1,1)$&$(8,5,2)$&$(9,5,1)$&$(8,6,1)$ \end{tabular} \end{center} Each is Clifford by the preceding subsections, Theorem~\ref{thm:D}, \cite[Proposition~2.1]{Nik2006} (an eigenvalue of multiplicity $9$), or \cite[Proposition~1]{Nik2005} for $(10,4,1)$ (in the notation of \cite[Proposition~1]{Nik2005}, $m=10$ and $\nu=5$ satisfy $16\ge3\nu$ and $16>(\nu+1)^2/4$). Each update has constant eigenvalues, so \cite[Lemma~3.2(1)]{Nik2006}, in the form of Remark~\ref{rem:lemma32-cl8} at the endpoint $(7,7,1)$, restores the original tensor. \subsection{The remaining patterns with multiplicities at most two} Shift the eigenvalue of multiplicity $M\in\{7,8\}$ to zero and let $r=15-M\le8$. By \cite[Lemma~3.1, p.~383]{Nik2006}, which applies because every nonzero eigenvalue has multiplicity one or two, there are $r$ orthogonal normalized fields for the nonzero eigenvalues. Divide each by the largest possible power of $q$ and homogenize all to the largest remaining odd degree $d$; then a component of a field of maximal degree is not divisible by $q$. With $P$ the $16\times r$ matrix of fields and $\Lambda$ the diagonal matrix of their eigenvalues, \[ P^TP=q^dI,\qquad P\Lambda P^T=q^{d-1}A,\qquad P\Lambda^2P^T=q^{d-2}A^2. \] If $d\ge3$, each diagonal entry of $P\Lambda^2P^T$ is a sum of $r\le8$ squares divisible by $q$, and \cite[Lemma~3.11]{Nik2006} makes every field component divisible by $q$, a contradiction. So $d=1$, the fields are $J_ix$ with orthogonal skew, pairwise anticommuting $J_i$, and the Jacobi formula is Clifford. (This is the proof of \cite[Lemma~3.3]{Nik2006} with at most eight columns instead of six.) This proves Theorem~\ref{thm:E}.