\documentclass[11pt,reqno]{amsart} \usepackage{amssymb} \usepackage[hmargin=1.10in,vmargin=1in]{geometry} \usepackage[hidelinks,pdfusetitle]{hyperref} \numberwithin{equation}{section} \newtheorem{theorem}{Theorem}[section] \newtheorem{proposition}[theorem]{Proposition} \newtheorem{lemma}[theorem]{Lemma} \theoremstyle{definition} \newtheorem{definition}[theorem]{Definition} \newcommand{\Gm}{\mathbb G_m} \newcommand{\GL}{\mathrm{GL}} \newcommand{\SL}{\mathrm{SL}} \newcommand{\QQ}{\mathbb Q} \newcommand{\ZZ}{\mathbb Z} \newcommand{\CC}{\mathbb C} \newcommand{\FF}{\mathbb F} \newcommand{\oo}{\mathfrak o} \newcommand{\Gg}{\mathfrak g} \DeclareMathOperator{\diag}{diag} \DeclareMathOperator{\Nm}{N} \DeclareMathOperator{\Hom}{Hom} \title[A counterexample to stability]{A counterexample to a stability question for unramified multiplicity one symmetric pairs} \author{Rui Chen} \hypersetup{pdftitle={A counterexample to stability question for unramified multiplicity one symmetric pairs}} \subjclass[2020]{Primary 22E50} \keywords{Symmetric varieties, unramified representations, multiplicity one} \begin{document} \begin{abstract} We construct a split reductive symmetric pair over $\QQ_{17}$ with multiplicity at most one for every irreducible unramified representation and which is not a stable symmetric pair. \end{abstract} \maketitle \section{Introduction} Let $k$ be a $p$-adic local field of characteristic zero with ring of integers $\oo$. For a symmetric pair $(G,H,\theta)$, fix a hyperspecial compact subgroup $K\subset G(k)$. An irreducible smooth complex representation $\pi$ is unramified with respect to $K$ if $\pi^K\ne0$. We say that the pair has \emph{unramified multiplicity one} if \[ \dim\Hom_{G(k)}\bigl(\pi,C^\infty(H(k)\backslash G(k))\bigr)\leq 1 \] for every such $\pi$. Here $G(k)$ acts by right translation. Aizenbud and Gourevitch call a symmetric pair satisfying Definition~\ref{stable} a good pair~\cite[Definition 7.1.6]{ag}. We use Carmeli's term \emph{stable} for the same property~\cite{carmeli}. We consider the following question. \begin{equation}\label{question} \text{Does unramified multiplicity one imply stability of a symmetric pair?} \end{equation} We take \[ k=\QQ_{17},\qquad (n_1,\ldots,n_6)=(4,4,8,8,4,8), \qquad h_j=n_j/2. \] Let $T=\Gm^6$, with coordinates $t=(s_1,t_1,s_2,t_2,u,v)$, and put \[ \tau(t)=(t_1,s_1,t_2,s_2,u,v). \] Write $p_i:T\to\Gm$ for the $i$-th coordinate projection. Define $\chi_j=c_j^{h_j}$ using the characters \begin{align} c_1&=p_1p_2p_3p_6, & c_2&=p_3, \label{eq:c}\\ c_3&=p_1p_6, & c_4&=p_1, \\ c_5&=p_1p_3p_5p_6, & c_6&=p_1p_3p_5. \label{eq:c6} \end{align} We introduce the group $G$ as \begin{equation}\label{eq:G} G=\left\{((A_j,B_j)_j,t)\in\prod_{j=1}^6\GL_{n_j}^2\times T\ \middle| \begin{aligned} \det A_j&=\chi_j(t),\\ \det B_j&=\chi_j(\tau t)\quad(1\leq j\leq6) \end{aligned}\right\}. \end{equation} The involution \begin{equation}\label{eq:H} \theta((A_j,B_j)_j,t)=((B_j,A_j)_j,\tau t) \end{equation} has fixed subgroup $H=G^{\theta}$. \begin{theorem}\label{thm:main} The symmetric pair defined by \eqref{eq:G}--\eqref{eq:H} has unramified multiplicity one (Proposition~\ref{prop:unramified-multiplicity-one}) and fails stability (Proposition~\ref{unstable}). \end{theorem} \section{Unramified multiplicity one}\label{sec:unramified} Write $S=T^\tau=\{(a,a,b,b,u,v)\}$ and let $p:G\to T$ be the torus projection. For a unit $x$ in a coefficient ring set $\delta_n(x)=\diag(x,1,\ldots,1)$. The homomorphism \[ \iota(t)=\bigl((\delta_{n_j}(\chi_j(t)), \delta_{n_j}(\chi_j(\tau t)))_j,t\bigr) \] splits $p$ over $\ZZ_{17}$ and restricts to a section $S\to H$. The kernels of $G\to T$ and $H\to S$ are respectively $\prod_j\SL_{n_j}^2$ and its blockwise diagonal copy $\prod_j\SL_{n_j}$. These descriptions give smooth group schemes with connected reductive geometric fibers. Indeed, the semidirect product descriptions give smoothness and connectedness, and in each fiber a connected normal unipotent subgroup maps trivially to the torus and hence lies in the semisimple kernel. The diagonal matrices give a split maximal torus in each group, since in each matrix factor the diagonal torus of $\SL_{n_j}$ already has diagonal centralizer. In particular $G$ is split reductive and $K=G(\ZZ_{17})$ is hyperspecial. For context we recall the general criterion of Sakellaridis. If $X$ is a homogeneous affine spherical variety for a connected split reductive group $L$, write $A^*$ for the unramified character torus of a maximal split torus of $L$, and use $A_X^*$, $\delta_{(X)}$, $W$, and $W_X$ for the corresponding objects of \cite[\S8]{sake}. The restriction map in that criterion is \[ \CC[A^*]^W\longrightarrow \CC[\delta_{(X)}^{1/2}A_X^*]^{W_X}. \] \begin{theorem}[{\cite[Corollary 8.0.4]{sake}}]\label{uniform} Assume the standing hypotheses of \cite[\S\S1.1, 1.7, 8]{sake}, including its integral orbit axioms and the single rational open Borel orbit hypothesis, and assume Statements 6.3.1 and 7.1.5 of loc.\ cit.\ for the spherical roots and virtual weighted colors. If the restriction map is surjective, then every irreducible unramified representation $\pi$ of $L(k)$ satisfies \[ \dim\Hom_{L(k)}\bigl(\pi,C^\infty(X(k))\bigr)\le 1 . \] \end{theorem} \begin{proof} Let $K_L$ be the hyperspecial subgroup and $\mathcal H$ its spherical Hecke algebra. By \cite[Corollary 8.0.4(1)]{sake}, $C_c^\infty(X(k))^{K_L}$ is cyclic over $\mathcal H$. Pair it with $C^\infty(X(k))^{K_L}$ by integration against an invariant measure. This pairing separates functions, since the compact open $K_L$-orbits have positive volume. If $f$ is a simultaneous eigenfunction with character $\omega$, its pairing with $h$ applied to a cyclic generator is its pairing with the generator multiplied by $\omega(h^\vee)$, where $h^\vee(g)=h(g^{-1})$. Thus each eigenspace has dimension at most one. An $L(k)$-equivariant map out of $\pi$ is determined by the image of a nonzero spherical vector, which spans $\pi^{K_L}$ and must map into the eigenspace for its Hecke character. \end{proof} For the pair \eqref{eq:G}--\eqref{eq:H} we give a direct proof. It rests on a choice of double-coset representatives for which averaging a spherical vector determines every $H(k)$-invariant functional from its value on that vector. \begin{lemma}\label{lem:compact} Every double coset in $H(k)\backslash G(k)/K$ has a representative $g_0$ such that $Kg_0K\subset H(k)g_0K$. \end{lemma} \begin{proof} Put $\varpi=17$ and let $\nu$ be the valuation with $\nu(\varpi)=1$. For $t=(s_1,t_1,s_2,t_2,u,v)$ the class of $t$ modulo $S(k)T(\oo)$ is determined by the two integers $x=\nu(t_1/s_1)$ and $y=\nu(t_2/s_2)$. Indeed, division by an element of $S(k)$ allows us to set $s_1,s_2,u,v$ equal to $1$, and $T(\oo)$ removes the remaining unit factors. Thus, using $\iota(S(k))$ on the left and $\iota(T(\oo))$ on the right, we may take the torus component to be \[ t=(1,\varpi^x,1,\varpi^y,1,\varpi^e), \qquad e=\begin{cases}1& x\ \text{odd and }y\ \text{even},\\ 0&\text{otherwise}.\end{cases} \] The last coordinate may be chosen freely because $S$ contains the last coordinate subtorus. The valuations of the six characters in \eqref{eq:c}--\eqref{eq:c6} are \begin{align*} (\nu c_j(t))_j &=(x+e,0,e,0,e,0),\\ (\nu c_j(\tau t))_j &=(x+y+e,y,x+e,x,x+y+e,x+y). \end{align*} In each column at least one entry is even. For $x$ even this holds throughout the first row. For $x,y$ both odd only its first entry is odd, and $x+y$ is even. In the remaining case $e=1$, the only odd entries of the first row are in columns $3$ and $5$, where the second row has even entries. Write the matrix components as $(A_j,B_j)$. Since $t$ has pure uniformizer powers as coordinates and $\chi_j=c_j^{n_j/2}$, at least one matrix in each pair has determinant $\lambda_j^{n_j}$ for some $\lambda_j\in k^\times$. If that matrix is $M_j$, multiplication on the left by $\lambda_j M_j^{-1}$ on both blocks is allowed in the diagonal $\SL_{n_j}(k)$ and makes $M_j$ scalar. We can also make the other block diagonal. Indeed, elementary divisors over $\oo$ give $P,Q\in\GL_{n_j}(\oo)$ with $PMQ$ diagonal for the other block $M$. Multiplying $P$ on the left and $Q$ on the right by diagonal unit matrices makes both determinants one and preserves diagonality. Apply $P$ on the left of both blocks, then $P^{-1}$ on the right of the scalar block and $Q$ on the right of the other. These operations belong to $H(k)$ and $K$, respectively. Denote the resulting representative by $g_0$. The subgroup $\iota(T(\oo))$ commutes with $g_0$, since all the matrices are diagonal. As $K=(\prod_j\SL_{n_j}(\oo)^2)\iota(T(\oo))$, it remains to move a pair $(U,V)\in\SL_{n_j}(\oo)^2$ past the $j$th block $(A,B)$. If $A$ is scalar, then \[ (UA,VB)=(V,V)(A,B)(V^{-1}U,I). \] If $B$ is scalar, then \[ (UA,VB)=(U,U)(A,B)(I,U^{-1}V). \] In each identity the left factor belongs to the diagonal subgroup and the right factor is integral of determinant one. Taking the product over $j$ proves the inclusion. \end{proof} \begin{proposition}\label{prop:unramified-multiplicity-one} The symmetric pair $(G,H)$ defined by \eqref{eq:G}--\eqref{eq:H} has unramified multiplicity one with respect to $K=G(\ZZ_{17})$. \end{proposition} \begin{proof} Let $\pi$ be irreducible and $K$-spherical. By the spherical representation theorem, $\pi^K=\CC w$ for any nonzero spherical vector $w$ \cite[\S6]{gross}. Fix $\ell\in\Hom_{H(k)}(\pi,\CC)$. For $g_0$ as in Lemma~\ref{lem:compact}, the function $g\mapsto\ell(\pi(g)w)$ is constant on $Kg_0K$, since it is left $H(k)$-invariant and right $K$-invariant. Give $K$ volume one. Then \[ \ell(\pi(g_0)w) =\ell\!\left(\int_K\pi(k'g_0)w\,dk'\right) =a_{g_0}\ell(w), \] where the averaged vector lies in $\pi^K$ and equals $a_{g_0}w$. The integral is a finite sum after passage to a sufficiently small open subgroup of $K$. If $\ell(w)=0$, the matrix coefficient therefore vanishes on every $H(k)g_0K$. The $G(k)$-translates of $w$ span $\pi$, so $\ell=0$. Evaluation at $w$ thus embeds $\Hom_{H(k)}(\pi,\CC)$ in $\CC$. Finally, evaluation of a $G(k)$-map at the identity coset identifies the Hom space in the introduction with $\Hom_{H(k)}(\pi,\CC)$. The inverse sends $\ell$ to $v\mapsto[H(k)g\mapsto\ell(\pi(g)v)]$. \end{proof} \subsection{Verification of Sakellaridis's criterion} \begin{proposition}\label{prop:sak-hypotheses} For $k=\QQ_{17}$ and the integral model of $X=H\backslash G$ over $\oo=\ZZ_{17}$, all the hypotheses of Theorem~\ref{uniform} hold. In particular $\delta_{(X)}=1$, and the restriction \[ \CC[A^*]^W\longrightarrow\CC[A_X^*]^{W_X} \] is surjective. Consequently Theorem~\ref{uniform} also gives Proposition~\ref{prop:unramified-multiplicity-one}. \end{proposition} \begin{proof} \emph{The open orbit and its integral model.} Take the Borel $B$ lower triangular in the first matrix of each pair and upper triangular in the second. Write $A$ for its diagonal torus. Identify $Q=T/S$ with $\Gm^2$ by $z(t)=(t_1/s_1,t_2/s_2)$. Define six characters of $Q$ by \[ (\eta_j(z))_j=(z_2,z_2,z_1,z_1,z_1z_2,z_1z_2). \] Indeed, $c_j(\tau t)/c_j(t)=\eta_j(z(t))$. Over $\oo$ the map $g\mapsto((A_j^{-1}B_j)_j,z(t))$ identifies $X$ with \begin{equation}\label{eq:integral-X} \left\{((Y_j)_j,z)\in\prod_j\GL_{n_j}\times Q \ \middle|\ \det Y_j=\eta_j(z)^{h_j}\right\}. \end{equation} To see this over any $\oo$-algebra, lift $z$ to $t=(1,z_1,1,z_2,1,1)$, set $A_j=\delta_{n_j}(\chi_j(t))$, and take $B_j=A_jY_j$. This gives a section, and two lifts differ by a unique left multiplier in $H$. In particular $X(k)=H(k)\backslash G(k)$. The model is smooth and affine, since multiplying $Y_j$ on the left by $\delta_{n_j}(\eta_j(z)^{-h_j})$ identifies it as a scheme with $Q\times\prod_j\SL_{n_j}$. The derived group of $G$ is $\prod_j\SL_{n_j}^2$ and is simply connected. The open orbit through $x_0=H$ is given by nonvanishing of all proper leading principal minors of the $Y_j$. Gaussian elimination factors each matrix there as lower unitriangular, diagonal, and upper unitriangular. The diagonal locus is $A_X=A/(A\cap H)$: the section just constructed takes diagonal tuples to $A$. Thus the open orbit is a product of $A_X$ with the two families of unitriangular groups over $\oo$. It is smooth, has homogeneous special fiber, and its $k$-points are one $B(k)$-orbit. The group $G$ and the stabilizer $H$ are unimodular, so $X(k)$ has an invariant measure. The open stabilizer $B\cap H=A\cap H$ is a split torus, and $\delta_B$ is trivial on it because the two triangular factors have opposite roots. This gives a $B(k)$-invariant measure on the open orbit. Every simple minimal parabolic enlarges that orbit, so $P(X)=B$, its Levi is $A$, and $\delta_{(X)}=1$. \emph{Bruhat strata and Axiom 2.2.1.} In \eqref{eq:integral-X} the $B$-strata are indexed by tuples $w=(w_j)\in\prod_j\mathfrak S_{n_j}$: the $j$th matrix lies in $B^-_{n_j}\dot w_j B^+_{n_j}$. Here $B^\pm_n$ are the triangular Borels of $\GL_n$, and we take signed permutation representatives of determinant one. The torus $Q$ is free to vary. Indeed, the diagonal factor used to trivialize \eqref{eq:integral-X} preserves these cells, the section $\iota$ moves between fibers, and the determinant-one triangular groups act transitively on each cell with fixed determinant. The integral cells are products of split tori and affine spaces by Bruhat elimination. Their frontiers are the usual unions of Bruhat cells, defined by the same northwest submatrix rank conditions over $\oo$ and in both fibers. In particular the boundaries, also for a union of two cells related by a minimal parabolic, acquire no extra special-fiber stratum: every boundary cell is flat over $\oo$ and its generic fiber is dense in it. To identify the rank-one quotient, take the representative $((\dot w_j)_j,1)$ of a stratum $Y$ and enlarge one triangular factor to $P_\alpha$. The stabilizer projected to $P_\alpha/\operatorname{rad}(P_\alpha)\simeq\mathrm{PGL}_2$ preserves one coordinate line: the other flag in the same block is a full coordinate flag in permutation order, and it singles out one of the two lines in the affected two-dimensional flag quotient. The image is its entire Borel. It contains the corresponding elementary unipotent matrices, acting on both flags with the permutation conjugacy, and the diagonal $\SL_2$-torus maps onto the adjoint torus as a group scheme for the faithfully flat topology. These matrix descriptions work over $\oo$. Thus $YP_\alpha/ \operatorname{rad}(P_\alpha)$ is a split Borel quotient of $\mathrm{PGL}_2$, of type U in \cite[\S2.2]{sake}. Together with the boundary description this gives Axiom 2.2.1 and excludes all type-N edges. \emph{The Cartan and Iwahori axioms.} Let $\Lambda=X_*(A)$ and $\Lambda_X=X_*(A_X)$. Write $\ell_j(\xi)=\langle c_j,\xi\rangle$ for $\xi\in X_*(T)$. An element of $\Lambda$ consists of an exponent vector $\xi$ and two integer vectors $a_j,b_j$ of length $n_j$ in each block with \[ \sum_i a_{ji}=h_j\ell_j(\xi),\qquad \sum_i b_{ji}=h_j\ell_j(\tau\xi). \] An element of $\Lambda_X$ consists of $q\in X_*(Q)$ and integer vectors $m_j$ with $\sum_i m_{ji}=h_j\langle\eta_j,q\rangle$. The projection of lattices is \begin{equation}\label{eq:lattice-restriction} \operatorname{pr}_X(a,b,\xi)=((b_j-a_j)_j,z_*(\xi)), \end{equation} where $z_*(\xi)$ denotes the image of $\xi$ in $X_*(Q)$. The antidominant cone consists of the $m_{j1}\le\cdots\le m_{jn_j}$ in each block. Indeed for a symmetric variety it is the image of the ambient antidominant cone over $\QQ$ \cite[\S2.1]{sake}. Here $a_j$ is decreasing and $b_j$ increasing in that cone. Conversely any increasing $m_j$ lifts rationally by taking $a_j$ constant with its prescribed sum and $b_j=a_j+m_j$. The invariants of a $K$-orbit are the two valuations of $z$ and the ordered elementary-divisor exponents of every $Y_j$. They determine the orbit. First normalize $z$ to pure powers of $17$ using $\iota(T(\oo))$. Smith normal form in block $j$ gives integral invertible $P,R$ with $PY_jR=D_j=\diag(17^{m_{ji}})_i$. At the normalized $z$ the determinants of $Y_j$ and $D_j$ agree. Thus $\det P\det R=1$, and replacing $(P,R)$ by $(EP,RE^{-1})$ for a diagonal unit matrix $E$ makes both determinants one while fixing $D_j$. These left and right operations on $Y_j$ are induced by the two $\SL_{n_j}(\oo)$ factors of $K$. Conversely the invariants are unchanged under $K$. Hence $X(k)/K$ is precisely the set of antidominant integral parameters just described (Axiom 2.4.1). For Axiom 2.4.2 take any diagonal point in $A_X^+$, with block $D=\diag(d_1,\ldots,d_n)$, so that $\nu(d_i)\le\nu(d_{i+1})$. For an element of the Iwahori $J$ write $(C,E)$ for its blocks. They are respectively lower and upper triangular modulo $17$. Gaussian elimination over $\oo$ gives $C^{-1}=L U$ with $L$ integral lower triangular invertible and $U$ upper unitriangular congruent to $I$ modulo $17$. Then $(D^{-1}UD)E$ is integral and upper triangular modulo $17$, so it has a factorization $L'R$ with $L'$ lower unitriangular and $R$ upper triangular integral invertible. Now \[ C^{-1}DE=L(DL'D^{-1})\,D\,R . \] The left factor is integral lower triangular invertible by the order of the valuations. Its determinant is $(\det C)^{-1}$, and $\det R=\det E$. Taking its inverse as the first block, $R$ as the second block, and the original torus component in $T(\oo)$ produces an element of $B(\oo)$ with exactly the same action on the given point. Thus $xJ\subset xB(\oo)$ for every $x\in A_X^+$. \emph{Spherical roots and virtual weighted colors.} On the Bruhat indices the ambient Weyl group acts separately by left and right permutations, since each simple step exchanges adjacent positions or adjacent values. The open index is the identity tuple. Its stabilizer is the blockwise diagonal $W_X=\prod_j\mathfrak S_{n_j}$, using $P(X)=B$ and \cite[\S2.2]{sake}. The simple reflection for positions $i,i+1$ in block $j$ is the product of the corresponding left and right simple reflections. Their length-two path passes through the same codimension-one cell reached from the open orbit on either side. Thus the normalized spherical root is the sum of the two corresponding simple roots, of type G in the notation of \cite[\S6.2]{sake}. Its supporting Levi has derived group $\SL_2\times\SL_2$. Thus \cite[Theorem 6.3.2]{sake} applies. In this case the rank-one computation is exactly the two-sided group computation in \cite[Example 5.4.1]{sake}: the adjacent root subgroups act by left and right multiplication on the same two-dimensional matrix block. This verifies Statement 6.3.1. For clarity, the color in this path is the divisor of the leading $i$-rowed principal minor $\Delta_{ji}$ in block $j$. Its valuation on $A_X$ is the cocharacter $\beta_{ji}=e_{ji}-e_{j,i+1}$, with zero torus component, where $e_{ji}$ denotes a standard coordinate vector in the $j$th block. Indeed the other leading minors and the $Q$-coordinates are units at its generic point, and $\Delta_{ji}$ vanishes there to order one. For the order of vanishing, take $z=(1,1)$ and the identity matrices with one two-by-two block replaced by $\left(\begin{smallmatrix}s&-1\\1&0\end{smallmatrix}\right)$ at positions $i,i+1$, where $\Delta_{ji}=s$ and all the other leading minors equal one. On the diagonal locus successive quotients of the leading minors are the diagonal coordinates. These evaluations give $\beta_{ji}$, and the Bruhat cells of codimension one exhaust the colors. Each color is defined over $k$, belongs to its single type-G root and has weight $\langle\beta_{ji},\rho_{P(X)}\rangle=1$, as seen by lifting $\beta_{ji}$ to the right simple coroot. The definition in \cite[\S7.1]{sake} therefore gives, with multiplicity one, \[ \Theta=\{(e_{ji}-e_{jl},+,1):1\le j\le6,\ 1\le i,l\le n_j,\ i\ne l\}. \] The cone of color valuations is generated by the adjacent differences $\beta_{ji}$, so the positive members have $i