\section{Further multiplicity reductions}\label{sec:smallroots} The classification also completes several polynomial-eigenfield reductions from \cite{Nik2006}. We give the additional arguments, including the exclusion needed when a linear perturbation mixes two eigenspaces. \begin{mainthm}\label{thm:E} An Osserman algebraic curvature tensor on $\R^{16}$ is Clifford if its reduced multiplicity pattern is one of \[ \begin{gathered} (7,6,1,1),\quad (8,5,1,1),\quad (7,5,2,1),\quad (8,4,2,1),\\ (7,5,1,1,1),\quad (8,4,1,1,1), \end{gathered} \] or if one root has multiplicity seven or eight and all the other roots have multiplicity at most two. The latter condition gives nine patterns: \[ \begin{array}{ll} (7,2,2,2,2),&(8,2,2,2,1),\\ (7,2,2,2,1,1),&(8,2,2,1,1,1),\\ (7,2,2,1,1,1,1),&(8,2,1,1,1,1,1),\\ (7,2,1,1,1,1,1,1),&(8,1,1,1,1,1,1,1),\\ (7,1,1,1,1,1,1,1,1).& \end{array} \] \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 \[ \norm{\pi_{N(x)}Kx}^2=c\norm x^2\qquad(x\ne0) \] for 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 $H_0,\ldots,H_8$ of Section~\ref{sec:Cayley}. Put $q=\norm x^2$, $h_i(x)=\ip{H_ix}x$ and $a_i(x)=\ip{H_ix}{Kx}$. The vectors $H_ix$ are orthogonal, have norm $\sqrt q$, and span $\R x+N(x)$. Since $K$ is skew, \[ \sum_i a_i(x)^2=cq^2,\qquad \sum_i h_i(x)^2=q^2. \] Differentiating the second equality in the direction $Kx$ gives $\sum_i h_i(x)a_i(x)=0$. The matrix of $a_i$ is the symmetric trace-free matrix $A_i=(H_iK-KH_i)/2$. If $c>0$, define \[ H_i(t)=\cos t\,H_i+\frac{\sin t}{\sqrt c}A_i. \] For every $t$, the associated quadratic map satisfies $\sum_i\ip{H_i(t)x}x^2=q^2$. This map cannot be constant on the unit sphere: otherwise $H_i(t)=b_iI$ for a unit vector $b$, whereas all its matrices have zero trace. Lemma~\ref{lem:threshold}(c) makes the $H_i(t)$ a symmetric Clifford system for every $t$. Its product $\Omega(t)=H_0(t)\cdots H_8(t)$ is a symmetric involution. Thus $\tr\Omega(t)$ is continuous and integer-valued, hence constant. But $\Omega(0)=\pm I$ on $\R^{16}$, while $\Omega(\pi)=-\Omega(0)$, a contradiction. If $c=0$, then $Kx\in L(x)$; skewness further gives $Kx\in E(x)$. Lemma~\ref{lem:nofield} gives $K=0$. \end{proof} \subsection{Polynomial tools and a three-column reduction} In this section $A(x)$ denotes the Jacobi matrix after shifting one reduced root to zero; it need not be the rank-seven normalization used earlier. Write $q=\norm x^2$ and $\mathcal K=\R[x_1,\ldots,x_{16}]/(q)$. We use three published algebraic inputs. By \cite[Theorem 3.1, p.~397]{Nik2006}, $\mathcal K$ is a UFD. By its Lemma 3.11, p.~398, if $q$ divides a sum of at most eight real polynomial squares, then it divides every polynomial being squared. Finally its Lemma 3.5, pp.~388--390, applies to a normalized cubic simple eigenfield $P$ satisfying \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} It says that a nonunit common divisor of the components of $P$ in $\mathcal K$ forces $P(x)=qJx$ for an orthogonal skew map $J$. Neither this statement nor its proof requires a root of multiplicity nine. The simple eigenfields of \cite[Lemma 3.1]{Nik2006} satisfy \eqref{eq:cubicidentities} after multiplication by $q$ if initially linear. We spell out the three-column step of \cite[pp.~387--389]{Nik2006}. Suppose $P_1,P_2,P_3$ are mutually orthogonal cubic eigenfields of squared norm $q^3$, and a weighted identity modulo $q$ has two positive and one negative coefficient. Rescale its columns by nonzero real constants and denote their images in $\mathcal K^{16}$ by $p_i$. Then \begin{equation}\label{eq:rankoneK} p_1p_1^T+p_2p_2^T=p_3p_3^T. \end{equation} Comparing diagonal and off-diagonal entries gives $(p_{1a}p_{2b}-p_{2a}p_{1b})^2=0$. Thus all these minors vanish. If any $p_i=0$, the original $P_i$ is $q$ times a linear eigenfield. Otherwise unique factorization gives a primitive vector $v$ and $p_1=u_1v$, $p_2=u_2v$. Equation \eqref{eq:rankoneK} gives $p_3=u_3v$ and $u_1^2+u_2^2=u_3^2$. Indeed a nonzero component of $v$ gives a square root of $u_1^2+u_2^2$ in the fraction field; it is integral over the integrally closed ring $\mathcal K$, hence belongs to $\mathcal K$. Cross entries fix the same sign in every component. Assume the negative column comes from a simple eigenvalue. If $u_3$ is a nonunit, Lemma 3.5 just quoted makes the original $P_3$ linear after division by $q$. If $u_3$ is a unit, homogeneity implies that $p_1,p_2$ are constant real multiples of $p_3$. Undoing the rescaling, a constant combination of the original fields therefore 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} where $J$ is 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 every fixed rank-one update by $C(Jx)(Jx)^T$ is still Osserman. \subsection{Two simple roots}\label{subsec:twosimple} Consider $(M,N,1,1)$ with $(M,N)=(7,6)$ or $(8,5)$. Shift the $M$-root to zero. Let $\lambda$ be the $N$-root and $\mu,\nu$ the simple roots. These values are distinct and nonzero. Lemma 3.1 of \cite{Nik2006} supplies cubic normalized eigenfields $P_1,P_2$, with linear fields homogenized by $q$. Their spectral projectors give \begin{equation}\label{eq:twocolumn} \mu(\mu-\lambda)P_1P_1^T+\nu(\nu-\lambda)P_2P_2^T =qA(A-\lambda qI). \end{equation} If the coefficients have the same sign, the sum-of-squares lemma applied to diagonal entries makes both fields $q$ times linear ones. Move one simple root into each large root. The result has pattern $(M+1,N+1)$, either $(8,7)$ or $(9,6)$. The former is classified by Theorem~\ref{thm:A}, with Cayley excluded by Lemma~\ref{lem:nofield}; the latter is Clifford by \cite[Proposition~2.1, p.~382]{Nik2006}. Restoring the two rank-one terms with \cite[Lemma 3.2(1)]{Nik2006} proves the claim. Otherwise interchange $\mu,\nu$ if necessary and put $a^2=\mu(\mu-\lambda)>0$, $b^2=-\nu(\nu-\lambda)>0$. In this subsection let $p_i$ denote the image of the unscaled $P_i$ in $\mathcal K^{16}$. Reducing \eqref{eq:twocolumn} modulo $q$ gives $(ap_1)(ap_1)^T=(bp_2)(bp_2)^T$. In a domain these vectors agree up to one global sign, by comparison of a nonzero component and then every cross entry. If both vanish the preceding linear-field argument applies. Otherwise change the sign of $P_2$ to obtain \[ aP_1-bP_2=\sqrt{a^2+b^2}\,qJx \] with $J$ orthogonal and skew. Set $C=\mu+\nu-\lambda\ne0$; nonvanishing follows from $a^2+b^2=(\mu-\nu)C$. On the unit sphere, subtracting $C(Jx)(Jx)^T$ changes the two-dimensional block to a fixed symmetric matrix of trace $\lambda$ and determinant \[ \mu\nu-C\,\frac{\nu a^2+\mu b^2}{a^2+b^2}=0, \qquad \nu a^2+\mu b^2=\mu\nu(\mu-\nu). \] The new roots are $0,\lambda$, so the pattern is $(M+1,N+1)$. 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}). Thus the endpoint is Clifford, and \cite[Lemma 3.2(1)]{Nik2006} restores the original tensor. \subsection{A double and a simple root} Consider $(M,N,2,1)$ with $(M,N)=(7,5)$ or $(8,4)$. Shift $M$ to zero, and call the other roots $\lambda,\mu,\nu$, where $\mu$ is double and $\nu$ is simple. There are cubic normalized fields $P_1,P_2$ for $\mu$ and $P_3$ for $\nu$, and \[ \mu(\mu-\lambda)(P_1P_1^T+P_2P_2^T) +\nu(\nu-\lambda)P_3P_3^T=qA(A-\lambda qI). \] For equal coefficient signs all fields are linear after division by $q$. For opposite signs multiply by $-1$ if necessary: the two double-root columns then have positive coefficients and the simple column has the negative coefficient. The three-column reduction applies. Its nonunit case gives a linear simple-root field; in its unit case \eqref{eq:linearcombination} is a linear double-root eigenfield. If a column vanishes modulo $q$, it already gives a linear eigenfield. Thus all cases produce one. Move that field into the zero root. The endpoints are \[ \begin{array}{c|cc} \text{initial}&\text{field from simple root}& \text{field from double root}\\ \hline (7,5,2,1)&(8,5,2)&(8,5,1,1)\\ (8,4,2,1)&(9,4,2)&(9,4,1,1). \end{array} \] These are Clifford by Theorem~\ref{thm:D}, the preceding subsection, or \cite[Proposition~2.1, p.~382]{Nik2006}. Restore the rank-one term by \cite[Lemma 3.2(1)]{Nik2006}. \subsection{Three simple roots} Consider $(M,N,1,1,1)$ with $(M,N)=(7,5)$ or $(8,4)$. There are five reduced roots. After shifting $M$ to zero, Lemma 3.1 of \cite{Nik2006} gives odd degree at most four, hence at most three. Let $\lambda$ be the $N$-root and $\mu_1,\mu_2,\mu_3$ the simple roots. The cubic spectral identity has weights $w_i=\mu_i(\mu_i-\lambda)$. If all signs agree, every field is linear after division by $q$. Otherwise choose the global sign and ordering so that two weights are positive and the third negative. The negative column is simple, so the three-column reduction applies. If it gives a linear eigenfield, move that field into the zero root, obtaining $(8,5,1,1)$ or $(9,4,1,1)$, which are Clifford by the Subsection~\ref{subsec:twosimple} and by \cite[Proposition~2.1]{Nik2006}; restore by \cite[Lemma~3.2(1)]{Nik2006}. Otherwise \eqref{eq:linearcombination} mixes $\mu_1,\mu_2$. Put $d=\mu_1c_2^2+\mu_2c_1^2$. The characteristic equation after subtracting $C(Jx)(Jx)^T$ is \[ (\mu_1-z)(\mu_2-z)-C(d-z)=0 . \] Choose $z\in\{0,\lambda\}$ with $d-z\ne0$; at least one choice works since $\lambda\ne0$. Then $C=(\mu_1-z)(\mu_2-z)/(d-z)\ne0$. The other root is the real constant $\theta=\mu_1+\mu_2-C-z$. The roots $0,\lambda,\mu_3$ survive unchanged and distinctly, so there is no two-root endpoint. The complete list of endpoints, including coincidences, is as follows. The column \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} Every entry is Clifford by the preceding subsections, Theorem~\ref{thm:D}, \cite[Proposition~2.1]{Nik2006} for the entries with a root of multiplicity $9$, or \cite[Proposition~1]{Nik2005} for $(10,4,1)$ (here $m=10$ 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} restores the original tensor. \subsection{All other multiplicities at most two} Finally shift the root of multiplicity $M\in\{7,8\}$ to zero. Let $r=15-M\le8$. The polynomial eigenframe lemma supplies $r$ orthogonal normalized fields for the nonzero roots. Divide each field by every common power of $q$ possible, then homogenize all fields to the largest remaining odd degree $d$. A component of a maximum-degree field is not divisible by $q$. Let $P$ be the $16\times r$ matrix of fields and let $\Lambda$ be the diagonal matrix of their nonzero eigenvalues. Then \[ P^TP=q^dI,\qquad P\Lambda P^T=q^{d-1}A,\qquad P\Lambda^2P^T=q^{d-2}A^2. \] If $d\ge3$, each diagonal entry on the last left side is a sum of $r\le8$ squares divisible by $q$. Lemma 3.11 of \cite{Nik2006} forces every field component to be divisible by $q$, a contradiction. Thus $d=1$. The fields are $J_ix$ with orthogonal skew, pairwise anticommuting $J_i$, giving a Clifford Jacobi formula. This is the proof of \cite[Lemma 3.3]{Nik2006} with its six columns replaced by at most eight. It completes Theorem~\ref{thm:E}.