# Real or imaginary: zeros of sector ground states of Weil-type quadratic forms
Draft 2026-11-01 (first version 07:50:32Z, revised 2026-11-01T12:53:41Z), span_of_dilates. Not yet checked by anyone else.
Tags: [U] unconditional proof; [num] computation, not used in any proof; [heur] heuristic.

**Abstract.** Connes and van Suijlekom proved that if the quadratic form of a real distribution on $[0,L]$ has a simple, isolated least eigenvalue with an even eigenfunction $\xi$, then the Fourier transform of $\xi$ has only real zeros. We record three complements. (1) The evenness hypothesis can be dropped. (2) If $\xi$ is only the least eigenfunction within one parity sector, the zeros of its transform lie on $\mathbb R\cup i\mathbb R$. The proof folds the sector matrix into a Loewner matrix on the nodes $n^2$. (3) The number of zeros on the positive imaginary axis is at most the number of sign changes of $\xi$ on the half window, by total positivity of $\sinh(xy)$ and $\cosh(xy)$. The finite lemma behind (1) and (2) is classical (boundary Nevanlinna–Pick interpolation); we claim the folding, the sector statement and (3). The odd-sector ground states of the Weil forms of $L(s,\chi_{-4})$ and $L(s,\chi_{-7})$ computed in rStation Paper 80 are sector minima, so (2) applies, and they are positive on the half window [num]; hence all zeros of their transforms are real. Positivity is not general: in 21 cases (11 characters, windows $L\ge2.4$) the ground state has up to two sign changes, yet all zeros are real; with one sign change, (3) and a parity count decide whether an imaginary pair exists. At the shorter window $L=1.6$ the odd ground state has a pair of imaginary zeros for every real odd character with $\chi(3)=1$, and it is the global ground state when $\chi(2)=\chi(3)=-1$, so that (1) is needed [num]. For $\zeta$ the odd ground states at eight windows with $1.2\le L\le4$ are positive, so all their zeros are real [num]. At two of these windows Osman had observed that the real zeros match the ordinates of $\zeta$. Nothing here bears on the Riemann hypothesis.

## 1. Setting and results

**Setting** (Connes–van Suijlekom [1, §4]). Let $L>0$ and let $D$ be a real distribution on $[0,L]$. On $H=L^2([0,L])$ take the orthonormal basis $U_n(x)=L^{-1/2}e^{2\pi inx/L}$, $n\in\mathbb Z$, and let $Q$ be the hermitian form [1, (6)] on trigonometric polynomials. By [1, Prop. 4.1–4.2] its matrix is
$$q_{nn}=a_n,\qquad q_{mn}=\frac{b_m-b_n}{m-n}\ (m\ne n),\qquad a_{-n}=a_n,\ b_{-n}=-b_n\in\mathbb R. \tag{1.1}$$
The reflection $\gamma f(x)=f(L-x)$ maps $U_n$ to $U_{-n}$, so $Q$ commutes with $\gamma$. Write $H_\pm=\{f:\gamma f=\pm f\}$ for the even and odd sectors, $E_N=\mathrm{span}\{U_n:|n|\le N\}$ and $E_N^\pm=E_N\cap H_\pm$. The Weil quadratic form of $\zeta$ restricted to test functions supported in a window of length $L$ is of this kind [1]. So is that of $L(s,\chi)$ for a real character $\chi$, because its explicit-formula distribution is real; we checked the resulting identity numerically on the matrices of Paper 80 (Section 5.3).

**Transform.** For $f\in H$ put $f_\sigma(t)=f(t+L/2)$ on $|t|\le L/2$, extended by $0$, and $\widehat{f_\sigma}(s)=\int f_\sigma(t)e^{-ist}\,dt$. The reflection $\gamma$ becomes $t\mapsto -t$. By [1, (20)], with $w=Ls/2\pi$ and $f=\sum_{|n|\le N}\xi_nU_n$,
$$\widehat{f_\sigma}(s)=\frac{L^{1/2}\sin\pi w}{\pi}\sum_{|n|\le N}\frac{\xi_n}{w-n}=c\,\frac{\sin\pi w}{\prod_{|j|\le N}(j-w)}\,P_\xi(w),\qquad P_\xi(w)=\sum_{k}\xi_k\prod_{j\ne k}(j-w), \tag{1.2}$$
with a constant $c\ne0$. The first factor is entire, and its zeros are the integers $|n|>N$. So the zeros of $\widehat{f_\sigma}$ are those integers together with the zeros of the polynomial $P_\xi$ of [1, (21)].

**Theorem 1** [U]. Assume that $Q$ defines a lower-bounded essentially self-adjoint operator and that the minimum of its spectrum is a simple isolated eigenvalue, with eigenfunction $\xi$. Then every zero of $\widehat{\xi_\sigma}$ is real.

Connes and van Suijlekom [1, Thm 6.1] prove this when $\xi$ is even. Since the eigenvalue is simple and $Q$ commutes with $\gamma$, $\xi$ is automatically even or odd, so Theorem 1 adds exactly the odd case. For Weil forms the even sector usually lies lower, as it does in 54 of the 55 cases with $L(s,\chi_D)$ where we compared the sectors (Sections 5.3–5.5) [num]. The exception shows that the odd case is not empty: on a window of length $L=1.6$ the odd ground state is the global one for $\chi_{-19}$, and hence for every $\chi_D$ with $D<0$ and $D\equiv5\pmod{24}$ (Section 5.5). For $\zeta$ on the same window (test functions supported in $[-0.8,0.8]$), by contrast, Zhu [8, Thm 6.2] certifies that the ground state is even, and he finds the even sector lower at every half-window in $[0.5,1.2]$ that he scanned [8, Table 1]. Theorem 2 still applies to the odd ground states of $\zeta$, and at the eight half-windows we computed, from $0.6$ to $2$, they have only real zeros (Section 5.6).

**Theorem 2** [U]. Assume that the restriction of $Q$ to the trigonometric polynomials in one sector $H_\pm$ defines a lower-bounded essentially self-adjoint operator on that sector, whose least spectral value is a simple isolated eigenvalue with eigenfunction $\xi$. Then every zero of $\widehat{\xi_\sigma}$ lies on $\mathbb R\cup i\mathbb R$. The same holds in finite dimension: if the least eigenvalue of $Q$ on $E_N^\pm$ is simple, its eigenvector's transform has all zeros on $\mathbb R\cup i\mathbb R$.

If $Q$ is essentially self-adjoint on all trigonometric polynomials, so is each sector restriction, because $(1\pm\gamma)/2$ maps trigonometric polynomials to trigonometric polynomials and commutes with $Q$. The point of Theorem 2 is that $\xi$ need not be a global minimiser. The alternative $i\mathbb R$ cannot be removed. It occurs in a random model (Section 5.1), and for Weil forms: at $L=1.6$ the odd-sector ground state of $L(s,\chi)$ has a pair of imaginary zeros for every real odd primitive character $\chi$ with $\chi(3)=1$ (Section 5.5) [num].

**Theorem 3** [U]. Let $f\in H_+$ or $f\in H_-$, $f\ne0$, and put $\ell=L/2$ and $g=f_\sigma$.
(i) The number of distinct zeros of $\widehat g$ on the open positive imaginary axis is at most the number of sign changes of $g$ on $(0,\ell)$, that is, of $f$ on $(L/2,L)$.
(ii) Suppose $g$ has exactly one sign change on $(0,\ell)$, and let $\varepsilon$ be the sign of $g$ after it. Put $m=\int_0^\ell t\,g(t)\,dt$ if $g$ is odd and $m=\int_0^\ell g(t)\,dt$ if $g$ is even, and assume $m\ne0$. Then $\widehat g$ has a zero on the open positive imaginary axis if and only if $\operatorname{sign}m\ne\varepsilon$, and there is then exactly one.

**Corollary 4** [U]. Let $\xi$ be as in Theorem 2, in finite or infinite dimension. If $\xi$ has no sign change on $(L/2,L)$, or exactly one with $\operatorname{sign}m=\varepsilon$ in the notation of Theorem 3(ii), then every zero of $\widehat{\xi_\sigma}$ is real.

**What is classical.** The finite-dimensional statement behind Theorems 1 and 2 (Lemma 5) says that a positive semidefinite Loewner matrix with a one-dimensional kernel defines a real-rooted pencil. It is the mechanism of the determinate case of boundary Nevanlinna–Pick interpolation, in which a singular positive Pick matrix forces a unique, real rational interpolant ([3, Thm 1.2 and §6] for one node; Sarason [4] for several). Our proof is the one of [1, Lemma 5.8], with the normalisation $\langle\eta|\xi\rangle=1$ replaced by a choice that exists for every kernel vector. Theorem 3(i) is an instance of the variation-diminishing property of totally positive kernels [5], and (ii) follows from (i) by a parity count. What we have not found in the literature is the folding of a sector to a Loewner matrix on $n^2$ (Lemma 6), the statement of Theorem 2, and the application of Theorem 3 to ground states. The question behind Theorems 1 and 2 has been asked numerically. Osman [10, §3] sampled random matrices of the form (1.1) and found $P_\xi$ real-rooted in 693 of 693 draws whose global ground state is odd, and in 686 of 1307 draws in which $\xi$ is the least odd vector and the global ground state is even. He records this as an observation, and reports that neither he nor Groskin knew of a real-rootedness theorem for odd ground vectors [10, §7]. Lemma 5, applied to $Q_N-\lambda$, proves the first observation. Theorem 2 in finite dimension says that in the second every non-real zero is purely imaginary, which we check under his sampling law (Section 5.1). Osman also computed the odd ground state of $\zeta$, which Groskin [9, §10] had not probed, and found that its real zeros match the ordinates of $\zeta$ [10, §2]; Section 5.6 shows that at his windows and ours it has no non-real zeros. Two preprints read the prime part of the Weil matrix as a Loewner divided-difference matrix. De Andrade Silva [11] gives the representation and a sign law for even and odd test vectors. Shi [6] builds a contrast pencil that requires a positive Gram matrix, which fails for odd ground vectors.

## 2. Finite dimension

Let $x_1,\dots,x_m$ be distinct reals and $X=\mathrm{diag}(x_j)$. A Loewner matrix is a real symmetric $A$ with $A_{jk}=(b_j-b_k)/(x_j-x_k)$ for $j\ne k$, for some real $b_j$, and an arbitrary real diagonal. Put $\eta=\sum_je_j$ and $\beta=\sum_jb_je_j$. Comparing entries,
$$XA-AX=|\beta\rangle\langle\eta|-|\eta\rangle\langle\beta|, \tag{2.1}$$
since both sides vanish on the diagonal. For $v\in\mathbb R^m$ and a fixed vector $\xi$ put $P_v(x)=\sum_jv_j\xi_j\prod_{i\ne j}(x_i-x)$.

**Lemma 5** [U]. Let $A\ge0$ be a Loewner matrix with $\ker A=\mathbb R\xi$. Then every $P_v$ with $v\in\mathrm{span}\{\eta,\beta\}$ that is not identically zero has only real zeros. In particular $P_\eta\not\equiv0$ is real-rooted, and $P_\eta$ and $P_\beta$ have interlacing zeros.

*Proof.* Put $a=\langle\eta|\xi\rangle$ and $b=\langle\beta|\xi\rangle$. By (2.1) and $A\xi=0$, $AX\xi=-a\beta+b\eta$. If $a=b=0$ then $X\xi\in\ker A$, so $X\xi=\mu\xi$; as the $x_j$ are distinct, $\xi$ is a multiple of some $e_k$, and then $a=\xi_k\ne0$, a contradiction. Hence the line $\mathcal V=\{v=p\beta+r\eta:\ pb+ra=1\}$ is not empty.

Take $v\in\mathcal V$, so $\langle v|\xi\rangle=1$, and put $X'=X-|X\xi\rangle\langle v|$. Then $X'\xi=0$. Write $u\wedge t=|u\rangle\langle t|-|t\rangle\langle u|$ and $w=AX\xi=-a\beta+b\eta$. By (2.1),
$$AX'-(AX')^{\mathsf T}=AX-XA-|w\rangle\langle v|+|v\rangle\langle w|=-\beta\wedge\eta+v\wedge w,$$
and $v\wedge w=(pb+ra)\,\beta\wedge\eta=\beta\wedge\eta$. So $AX'$ is symmetric: $X'$ is self-adjoint for the semi-inner product $\langle f|Ag\rangle$. Since $X'$ maps $\ker A$ to $0$, it induces a self-adjoint operator on the Euclidean space $\mathbb R^m/\ker A$, with real eigenvalues $\mu_1,\dots,\mu_{m-1}$, and $\det(X'-x)=-x\prod_k(\mu_k-x)$. By the matrix determinant lemma and $\langle v|\xi\rangle=1$,
$$\det(X'-x)=\det(X-x)\Big(1-\sum_j\frac{v_j\xi_jx_j}{x_j-x}\Big)=\det(X-x)\sum_j\frac{-x\,v_j\xi_j}{x_j-x}=-x\,P_v(x).$$
So $P_v(x)=\prod_k(\mu_k-x)$ is real-rooted for every $v\in\mathcal V$. The only direction of $\mathrm{span}\{\eta,\beta\}$ that $\mathcal V$ misses is $v_0=a\beta-b\eta$, with $\langle v_0|\xi\rangle=0$. Fix $v_1\in\mathcal V$. For $t\ne0$, $v_0+tv_1=t(v_1+v_0/t)$ with $v_1+v_0/t\in\mathcal V$, so $P_{v_0+tv_1}$ is real-rooted, and it converges to $P_{v_0}$ coefficientwise as $t\to0$. If $P_{v_0}\not\equiv0$ had a non-real zero, Hurwitz's theorem on a small disc around it, disjoint from $\mathbb R$, would give non-real zeros of $P_{v_0+tv_1}$ for small $t$. Finally $P_\eta\not\equiv0$, because its coefficients in the basis $\prod_{i\ne j}(x_i-x)$ are the $\xi_j$. Real-rootedness of the whole pencil is equivalent to interlacing (Hermite–Kakeya–Obreschkoff). $\square$

The quotient $P_\beta/P_\eta$ takes the value $b_j$ at $x_j$ (when $\xi_j\ne0$), so it is a rational interpolant of the data $(x_j,b_j)$; this is the boundary Nevanlinna–Pick reading of Lemma 5 mentioned in Section 1. For (1.1) with $x_n=n$, Lemma 5 applied to $Q_N-\lambda$ is [1, Thm 5.6] without the hypothesis that $\xi$ is even; the normalisation $\langle\eta|\xi\rangle=1$ used in [1, Lemma 5.8] is the case $p=0$.

**Lemma 6** (folding) [U]. Let $Q$ be as in (1.1) on $E_N$.
(i) In the basis $o^-_j=(U_j-U_{-j})/\sqrt2$, $1\le j\le N$, of $E_N^-$, $Q(o^-_j,o^-_k)=J_jJ_k\Lambda^-_{jk}$ with $J_j=\sqrt2\,j$, where $\Lambda^-$ is the Loewner matrix of $\varphi_j=b_j/j$ on the nodes $j^2$, with diagonal $(a_j-b_j/j)/2j^2$.
(ii) In the basis $o^+_0=U_0$, $o^+_j=(U_j+U_{-j})/\sqrt2$ of $E_N^+$, $Q(o^+_j,o^+_k)=J_jJ_k\Lambda^+_{jk}$ with $J_0=1$, $J_j=\sqrt2$ ($j\ge1$), where $\Lambda^+$ is the Loewner matrix of $\kappa_j=jb_j$ on the nodes $j^2$, $0\le j\le N$, with diagonal $a_0$ and $(a_j+b_j/j)/2$.
(iii) For $f=\sum_jy_jo^\pm_j$, with $W(x)=\sum_j(Jy)_j/(x-j^2)$: in the odd case $\sum_n\xi_n/(w-n)=W(w^2)$, and in the even case $\sum_n\xi_n/(w-n)=w\,W(w^2)$.

*Proof.* (i) By $q_{-m,-n}=q_{mn}$, $Q(o^-_j,o^-_k)=q_{jk}-q_{j,-k}$. For $j\ne k$, using $b_{-k}=-b_k$,
$$\frac{b_j-b_k}{j-k}-\frac{b_j+b_k}{j+k}=\frac{2(kb_j-jb_k)}{j^2-k^2}=2jk\,\frac{b_j/j-b_k/k}{j^2-k^2}.$$
For $j=k$ it is $q_{jj}-q_{j,-j}=a_j-b_j/j$. (ii) For $j\ne k$, both $\ge1$, $q_{jk}+q_{j,-k}=2(jb_j-kb_k)/(j^2-k^2)$. For $k=0$, $b_0=0$ and $Q(o^+_j,o^+_0)=\sqrt2\,q_{j0}=\sqrt2\,b_j/j=J_jJ_0\,(\kappa_j-\kappa_0)/(j^2-0)$. The diagonal is $a_j+b_j/j$ for $j\ge1$ and $a_0$ for $j=0$. (iii) $\frac1{w-j}\mp\frac1{w+j}$ equals $\frac{2j}{w^2-j^2}$ or $\frac{2w}{w^2-j^2}$. $\square$

A shift $Q-\lambda$ changes only the diagonal: $Q^\pm-\lambda=J(\Lambda^\pm-\lambda J^{-2})J$, and $\Lambda^\pm-\lambda J^{-2}$ is again a Loewner matrix on $j^2$.

**Proposition 7** (Theorem 2 in finite dimension) [U]. Let $\lambda$ be the least eigenvalue of $Q$ on $E_N^\pm$, simple there, with eigenvector $y$. Then $W$ has only real zeros, and every zero of $\widehat{f_\sigma}$, $f=\sum y_jo^\pm_j$, lies on $\mathbb R\cup i\mathbb R$.

*Proof.* $\Lambda^\pm-\lambda J^{-2}=J^{-1}(Q^\pm-\lambda)J^{-1}\ge0$ has kernel $\mathbb R\,Jy$. Lemma 5 on the nodes $j^2$ gives that $R(x)=\sum_j(Jy)_j\prod_{i\ne j}(i^2-x)=-W(x)\prod_i(i^2-x)$ is real-rooted. By (1.2) and Lemma 6(iii), in the odd case $\widehat{f_\sigma}(s)=\pi^{-1}L^{1/2}\sin(\pi w)\,W(w^2)=-\pi^{-1}L^{1/2}\,\frac{\sin\pi w}{\prod_{1\le i\le N}(i^2-w^2)}\,R(w^2)$, and the fraction is entire with zeros at $w=0$ and the integers $|n|>N$. So the remaining zeros are the $w=\pm\sqrt x$ with $R(x)=0$, which are real if $x\ge0$ and purely imaginary if $x<0$. In the even case $\widehat{f_\sigma}(s)=\pi^{-1}L^{1/2}\,w\sin(\pi w)\,W(w^2)$, the product runs over $0\le i\le N$, and $w\sin(\pi w)/\prod_{0\le i\le N}(i^2-w^2)$ is entire with zeros only at the integers $|n|>N$. $\square$

## 3. Infinite dimension

*Proof of Theorems 1 and 2.* We follow the proof of [1, Thm 6.1] and only list what changes. For Theorem 2 work in the closed invariant subspace $H_\pm$ with the restricted operator $A_\pm$; the trigonometric polynomials of $H_\pm$ are a core by hypothesis. Given $\epsilon>0$ choose a trigonometric polynomial $\eta_\epsilon\in H_\pm$ with $\|\eta_\epsilon\|=1$, $\|\xi-\eta_\epsilon\|<\epsilon$ and $Q(\eta_\epsilon)<\lambda+\epsilon$. The spectrum of $A_\pm$ lies in $\{\lambda\}\cup[\lambda+\delta,\infty)$, so $Q(\alpha)\ge(\lambda+\delta)\|\alpha\|^2$ for $\alpha\in H_\pm$, $\alpha\perp\xi$. The min–max step of [1] then shows the following when $\epsilon<\delta/2$ and $N$ is at least the degree of $\eta_\epsilon$: the least eigenvalue $\lambda_N$ of $Q$ on $E_N^\pm$ is simple, and its eigenvector $\xi_N$ satisfies $\|\xi-\xi_N\|\le\epsilon+\sqrt{\epsilon/\delta}$. Every vector used lies in $H_\pm$. In [1] the step "$\xi_N$ is even as well" is needed to apply [1, Thm 5.6]; here Proposition 7 applies to $\xi_N$ directly, so the zeros of $\widehat{(\xi_N)_\sigma}$ lie on $\mathbb R\cup i\mathbb R$. As $\xi_N\to\xi$ in $L^2$ on a fixed window, $\widehat{(\xi_N)_\sigma}\to\widehat{\xi_\sigma}$ uniformly on compact sets, and $\widehat{\xi_\sigma}\not\equiv0$. The set $\mathbb R\cup i\mathbb R$ is closed, so Hurwitz's theorem gives Theorem 2. For Theorem 1 run the same argument in $H$, with Lemma 5 applied to $Q_N-\lambda_N$ on the nodes $n\in\{-N,\dots,N\}$ in place of [1, Thm 5.6]; the parity of $\xi_N$ is not used, and the zeros of $P_{\xi_N}$, hence of $\widehat{(\xi_N)_\sigma}$, are real. $\square$

## 4. Imaginary zeros and sign changes

*Proof of Theorem 3.* (i) The function $g$ is odd or even on $[-\ell,\ell]$. For $y>0$,
$$\widehat g(iy)=\int_{-\ell}^{\ell}g(t)e^{yt}\,dt=2\int_0^\ell g(t)\sinh(yt)\,dt\quad\text{or}\quad 2\int_0^\ell g(t)\cosh(yt)\,dt.$$
The kernels $\sinh(ty)=\sum_k(ty)^{2k+1}/(2k+1)!$ and $\cosh(ty)=\sum_k(ty)^{2k}/(2k)!$ are power series in $ty$ with positive coefficients and strictly increasing exponents. By the Cauchy–Binet formula, each minor $\det[K(t_i,y_j)]$ with $0<t_1<\dots<t_r$ and $0<y_1<\dots<y_r$ is a sum of products of positive coefficients and two generalized Vandermonde determinants $\det[t_i^{n_k}]\det[y_j^{n_k}]$, which are positive. So both kernels are strictly totally positive on $(0,\infty)^2$, and the variation-diminishing property of strictly totally positive kernels [5, Ch. 5], in the form $S^+(\widehat g(i\,\cdot))\le S^-(g)$, bounds the number of zeros of $y\mapsto\widehat g(iy)$ on $(0,\infty)$ by the number of sign changes of $g$ on $(0,\ell)$. Here $S^-$ counts sign changes and $S^+$ counts sign changes with zeros given either sign, so $S^+$ is at least the number of distinct zeros. The function $\widehat g(iy)$ is not identically zero, since $\widehat g$ is entire and $g\ne0$. In terms of $f$, $g$ on $(0,\ell)$ is $f$ on $(L/2,L)$.

(ii) Put $h(y)=\widehat g(iy)$ and let $t_0$ be the sign change, so $\varepsilon g\ge0$ on $(t_0,\ell)$ and $\varepsilon g\le0$ on $(0,t_0)$. As $y\to0^+$, $h(y)/2y\to m$ in the odd case and $h(y)\to2m$ in the even case, so $h$ has the sign of $m$ near $0$. Choose $t_1\in(t_0,\ell)$ with $\int_{t_1}^\ell|g|>0$. In the odd case $\varepsilon h(y)\ge2\sinh(yt_1)\int_{t_1}^\ell|g|-2\sinh(yt_0)\int_0^{t_0}|g|$, which is positive for large $y$ because $\sinh(yt_0)/\sinh(yt_1)\to0$; the even case is the same with $\cosh$. So $h$ has the sign $\varepsilon$ for large $y$. By (i) in the form $S^+(h)\le1$, $h$ has at most one zero. If $\operatorname{sign}m\ne\varepsilon$ it has one. If $\operatorname{sign}m=\varepsilon$, a zero $y_1$ would give $S^+(h)\ge2$: take points $y_a<y_1<y_b$ where $h$ has the sign $\varepsilon$, and give $y_1$ the sign $-\varepsilon$. $\square$

If $g$ has no sign change on $(0,\ell)$ the bound needs no theory: $\widehat g(iy)$ is then an integral of a one-signed function against a positive kernel. Corollary 4 follows from Theorem 2 or Proposition 7 together with Theorem 3(i) or (ii), because $\widehat g(-z)=\pm\widehat g(z)$, so the zeros on the negative imaginary axis mirror those on the positive one. With two or more sign changes, part (i) bounds but does not count: the second eigenvectors in Section 5.3 have two and three sign changes, and $W(-v^2)$ has no sign change on a grid of $0<v\le5K$ [num].

*How imaginary pairs appear* [U]. Let the matrix of $Q$ on $E_N^-$ depend continuously on a parameter, with a simple least eigenvalue throughout, and choose the eigenvector $y$ continuously. By Proposition 7 the zeros of $\widehat{g}$ other than the integers $|n|>N$ are the $\pm\sqrt x$ for the zeros $x$ of the real-rooted polynomial $R$, which has degree at most $N-1$ and leading coefficient $(-1)^{N-1}\sum_j(Jy)_j$. A zero of $R$ can change sign only by passing through $x=0$, where $R(0)=-W(0)\prod_ii^2$ vanishes, or by passing through infinity, where $\sum_j(Jy)_j=0$. Since $W(0)$ is a fixed multiple of $m$ and $\sum_j(Jy)_j$ a fixed multiple of $g'(\ell)$, a pair of imaginary zeros is born or dies only at the origin, when $m$ changes sign, or at infinity, when $g'(\ell)$ changes sign, that is, when a sign change of $g$ enters or leaves through the end of the half window.

## 5. Computations [num]

All predictions below were registered before the computation. The scripts are in the supplement; they use mpmath, with the stated number of digits, and none of them is interval arithmetic.

**5.1 Random matrices.** Take (1.1) with $b_j\sim N(0,1)$ for $j\ge1$, $b_{-j}=-b_j$, $b_0=0$, and $a_j=a_{-j}\sim N(0,4)$, for $N=3,5,8$ (60, 60 and 30 seeds; 60 digits).
- Global minimum (Lemma 5 on $Q_N-\lambda$). It was odd in 52 instances and even in 98. $P_\xi$ was real-rooted, with zeros interlacing those of $P_\beta$, in 52 of 52 and 98 of 98.
- Sector minima that are not global minima (Proposition 7). For the least odd vector when the global minimum is even, all zeros lay on $\mathbb R\cup i\mathbb R$ in 98 of 98, and 43 of the 98 had imaginary zeros. For the least even vector when the global minimum is odd: 52 of 52 on $\mathbb R\cup i\mathbb R$, all 52 with imaginary zeros.
- Control. The second odd eigenvector had zeros off both axes in 72 of 150 instances, so the detector does find such zeros.
- Osman's sampling law [10, §3]: $a_0,\dots,a_N$ and $b_1,\dots,b_N$ i.i.d. $N(0,1)$, $N=3,4,5,6,8$, 300 draws each. In the 1001 draws with an even global minimum, the least odd vector had all zeros on $\mathbb R\cup i\mathbb R$; 496 had imaginary zeros, with $|\mathrm{Im}\,w|$ up to 116 and real parts below $10^{-57}$ relative to $|w|$. His non-real-rooted fraction was 621 of 1307. All 499 draws with an odd global minimum were real-rooted (his 693 of 693).
- Without any symmetry ($a,b$ arbitrary; 80 instances at $N=3,5$), the least eigenvector's pencil was real-rooted and interlacing in 80 of 80, and so was the limit direction $v_0$ of the proof of Lemma 5.

**5.2 Theorem 3.** For 80 random odd vectors ($K=8,12$ sine modes on $[-2,2]$, Gaussian coefficients scaled by $j^{-s}$, $s\in\{0,1,2\}$), the number of imaginary zero pairs never exceeded the number of sign changes on $(0,2)$ (5000-point grid). 36 of the 80 had imaginary zeros, and the bound was attained in 27.

**5.3 The ground states of Paper 80.** Paper 80 [2] computes the least eigenvector of the Weil form of $L(s,\chi_D)$ on odd test functions supported in $[-\ell,\ell]$, in the span of $\sin(j\pi t/\ell)$, $1\le j\le K$. This is $E_K^-$ for $L=2\ell$; the basis differs from $o^-_j$ by the phases $i(-1)^j$, which is why the identity of Lemma 6(i) holds for the matrix of [2] only after the sign change $(-1)^{j+k}$. We checked that identity on the matrix of [2] for $\chi_{-4}$, $\ell=1.6$, $K=20$: after the sign change, $r_{jk}=Q_{jk}(j^2-k^2)/2jk$ satisfies $r_{jk}+r_{kl}=r_{jl}$ for all distinct $j,k,l$ to $1\times10^{-40}$ relative, at 40 digits. (Our first registered check omitted the sign and failed.)

We used the two saved ground states of [2]: $\chi_{-4}$ with $\ell=2.2$, $K=380$ (170 digits), and $\chi_{-7}$ with $\ell=2.4$, $K=320$ (140 digits). Their least eigenvalues are simple: $\mu_1=3.47\times10^{-101}$, $\mu_2=3.87\times10^{-92}$ for $\chi_{-4}$ and $\mu_1=3.48\times10^{-85}$, $\mu_2=2.01\times10^{-76}$ for $\chi_{-7}$. Proposition 7 applies to them whether or not they are global minima. We did not compute the even sectors at these windows. For $\chi_{-4}$ at $\ell=1.8$, $K=160$, the least even eigenvalue is $1.70\times10^{-45}$ against $2.79\times10^{-41}$ for the odd one, so there the odd ground state is not a global minimum and Theorem 1 does not reach it.
- *Sign.* On 19999 equally spaced points of $(0,\ell)$, both ground states are positive. The maxima are 1.51 and 1.37 and the minima $6\times10^{-50}$ and $5\times10^{-42}$, attained next to $t=\ell$. The functions are below $10^{-20}$ on the last 9% and 6% of the half window, and their derivative at $\ell$ is proportional to $\sum_j(Jy)_j$, which is $3\times10^{-46}$ and $2\times10^{-38}$ of the largest weight $|(Jy)_j|$. There the sign rests on cancellation. The working precision resolves it, and the eigenvector error, about the rounding level divided by $\mu_2-\mu_1$, is near $10^{-79}$ and $10^{-64}$, far below these minima [heur]. The second eigenvectors change sign 2 and 3 times. By Proposition 7 and Theorem 3, every zero of the two transforms is real. The grid does not see the last interval $(\ell-\ell/20000,\ell)$, and that is the part of $(0,\ell)$ that decides $\widehat g(iy)$ for large $y$. The independent count below does not depend on the grid in $t$ and covers it.
- *Independent count.* The function $W$ has exactly $K-1$ zeros, since its numerator $R$ has degree $K-1$ with leading coefficient $\pm\sum_j(Jy)_j\ne0$. Counting sign changes of $W$ on $(0,\infty)$ locates all of them: 379 of 379 and 319 of 319. We sampled 64 equally spaced points in each gap between nodes, geometric points near each node, and a geometric grid up to about $10^6$ times the last node. With equally spaced points alone the count was 377 and 317; in both cases the two missing zeros were beyond the last node, where 5 zeros lie. None lies below the first node. This count uses neither Proposition 7 nor Theorem 3.
- *Agreement with [2].* On $(0,T_{\max}]$ ($T_{\max}=360$ and $300$) the count gives 244 and 222 zeros, gap by gap the same as the root lists saved by [2]. Those lists contain one more zero each, just above $T_{\max}$.

Both computations are on grids, so a close pair of sign changes could be missed. An interval-arithmetic version of the sign check is the obvious next step and is not done here.

**5.4 Other characters and windows.** We repeated the computations with the matrices of [2] for $\chi_D$, $D\in\{-3,-4,-7,-8,-11\}$ and $\ell\in\{1.2,1.6,2.0\}$, and for six more characters at $\ell=2.0$, always with $K=100$. That makes 21 cases. The least odd eigenvalue is simple in each ($\mu_2/\mu_1\ge9.3$) and lies between $0.17$ ($\chi_{-11}$, $\ell=1.2$) and $1.6\times10^{-84}$ ($\chi_{-3}$, $\ell=2.0$). We used 120 digits. The 120-digit eigenvector error, about $10^{-120}/\mu_2$, is at most $5\times10^{-45}$ ($\chi_{-3}$, $\ell=2.0$). We repeated the sign check at 200 digits in the seven cases with $\mu_2<10^{-10}$, and the count below in the five of them with no sign change. The sign counts at 120 and 200 digits agree. The largest difference between the two on the grid is $1.7\times10^{-45}$ ($\chi_{-3}$, $\ell=2.0$), against a smallest grid value of $3.8\times10^{-41}$.
- *Sign* (4000-point grid on $(0,\ell)$). The ground state has no sign change for $\chi_{-3}$, $\chi_{-4}$ and $\chi_{-7}$ at all three windows and for $\chi_{-15}$ and $\chi_{-23}$. It has one for $\chi_{-8}$ and $\chi_{-11}$ at all three windows and for $\chi_{-19}$, $\chi_{-24}$ and $\chi_{-31}$, and two for $\chi_{-20}$. We had predicted positivity in all of the first 15 cases. A rule we then guessed, that the ground state changes sign exactly when $\chi(3)=1$, failed on the next six.
- *Where the sign changes.* For $\chi_{-8}$ the sign change lies at $t_0=0.40,0.33,0.30$ and for $\chi_{-11}$ at $t_0=0.70,0.65,0.65$ ($\ell=1.2,1.6,2.0$), so it hardly moves with $\ell$. For $\chi_{-8}$ the lobe before $t_0$ is small (maximum 0.26, 0.14, 0.11 against 1.19, 1.32, 1.45 after it). For $\chi_{-11}$ the two lobes are comparable (0.84 against 1.14 at $\ell=1.2$). For $\chi_{-20}$ the sign changes are at $t/\ell=0.229$ and $0.395$, with lobes of maximum 0.20, 0.17 and 1.47.
- *Theorem 3(ii).* Up to one common factor, $W(0)$ has the sign of $m$ and $W$ near $-\infty$ the sign of $\varepsilon$, because $\widehat g(iy)$ is a fixed multiple of $\sinh(\ell y)\,W(-\ell^2y^2/\pi^2)$ by (1.2) and Lemma 6(iii). In the nine cases with one sign change the two signs agree, so there is no imaginary zero.
- *Independent count.* In all 21 cases, counting sign changes of $W$ as in Section 5.3 finds 99 zeros on $(0,\infty)$, which is all of them, and none on a geometric grid of $(-\infty,0)$. This also covers the last grid interval in $t$, which the sign check cannot see (Section 5.3).
- *Even sector.* We computed the least even eigenvalue in all 21 cases. It lies lower in every case, by factors between 29 ($\chi_{-11}$, $\ell=1.2$) and $2.7\times10^4$ ($\chi_{-4}$, $\ell=2.0$), so none of these odd ground states is a global minimum, and Theorem 1 does not reach them.

So in 21 cases over 11 characters the odd-sector ground state has 0, 1 or 2 sign changes (11, 9 and 1 cases), and its transform has only real zeros in every case [num]. The imaginary alternative of Theorem 2 did not occur at these windows. It does occur at shorter ones (Section 5.5).

**5.5 Short windows.** For the same 11 characters at $\ell\in\{0.6,0.8,1.0\}$ (33 cases; $K=60$, 60 digits) the prime sum contains only $n\le e^{2\ell}\le7.4$. The conductor enters the form only through $\log q$ times $\|f\|^2$, a multiple of the identity, so the eigenvectors depend on $\chi$ only through $\chi(n)$ for these $n$. For example, $\chi_{-8}$ and $\chi_{-20}$ have the same eigenvectors at $\ell=0.8$, and their eigenvalues differ by $\log(20/8)$. At $\ell=0.6$ and $0.8$ only $\chi(2)$ and $\chi(3)$ matter, since $\chi(4)=\chi(2)^2$, and our 11 characters realise all nine pairs $(\chi(2),\chi(3))\in\{-1,0,1\}^2$. So at these two windows the findings below hold for every real odd primitive character, at $K=60$.
- *Imaginary zeros.* In 8 of the 33 cases $W$ has one negative zero, so the transform of the odd ground state has a pair of imaginary zeros $\pm iy$: $\chi_{-23}$ at $\ell=0.6$ ($y=2.27$); $\chi_{-8}$, $\chi_{-20}$ ($y=5.33$), $\chi_{-11}$ ($7.80$) and $\chi_{-23}$ ($9.67$) at $\ell=0.8$; $\chi_{-19}$ ($9.06$), $\chi_{-24}$ ($24.2$) and $\chi_{-31}$ ($2.16$) at $\ell=1.0$. Six of these ground states have one sign change and two have two. In Sections 5.4 and 5.5 together, 24 ground states have exactly one sign change, at $t_0$. A pair occurs in the 6 with $t_0/\ell$ between $0.63$ and $0.94$, where the lobe next to $t=\ell$ is the smaller one, and in none of the 18 with $t_0/\ell$ between $0.15$ and $0.72$. The two ranges overlap, so $t_0$ alone does not decide; by Theorem 3(ii) the sign of $m$ does. At $\ell=0.8$ a pair occurs exactly when $\chi(3)=1$, and at $\ell=0.6$ exactly when $\chi(2)=\chi(3)=1$. In all 33 cases the zeros of $W$ on the two half-lines add up to $59=K-1$, so all other zeros are real, as Proposition 7 requires.
- *Dependence on $K$.* We recomputed six of the pairs at $K=100$ and $140$ (80 digits). All persist. The values of $y$ at $K=60,100,140$ are $9.058, 8.970, 8.932$ ($\chi_{-19}$, $\ell=1.0$), $5.329, 5.247, 5.213$ ($\chi_{-8}$, $\ell=0.8$), $7.800, 7.711, 7.675$ ($\chi_{-11}$, $\ell=0.8$), $9.672, 9.472, 9.370$ ($\chi_{-23}$, $\ell=0.8$), $2.267, 2.231, 2.216$ ($\chi_{-23}$, $\ell=0.6$) and $24.19, 23.55, 23.29$ ($\chi_{-24}$, $\ell=1.0$). So each of the three classes with $\chi(3)=1$ keeps its pair at $\ell=0.8$ up to $K=140$. The values decrease more slowly as $K$ grows, so we expect a limit near them, but we have not proved that the pairs survive as $K\to\infty$. For $\chi_{-23}$ at $\ell=0.8$ the two least odd eigenvalues are close: $\mu_1=3.13550$ and $\mu_2-\mu_1=8.4\times10^{-5}$ at $K=60$, $8.3\times10^{-5}$ at $K=100$. The gap is far above the working precision, so the eigenvalue is simple and Proposition 7 applies; the ground state keeps its two sign changes at all three $K$.
- *An odd global ground state.* For $\chi_{-19}$ at $\ell=0.8$ the least odd eigenvalue lies below the least even one: $1.4267$ against $1.7503$ at $K=60$, and $1.4258$ against $1.7502$ at $K=140$. So this odd ground state is the global one, and Theorem 1, in the case that [1] does not cover, applies to it. It has no sign change and all its zeros are real. In the other 32 cases the even sector lies lower. By the remark at the start of this section the same holds for every $\chi_D$ with $\chi_D(2)=\chi_D(3)=-1$, that is $D\equiv5\pmod{24}$; among our 11 characters only $\chi_{-19}$ is in this class. As a check, the eigenvalues for $\chi_{-43}$ are those for $\chi_{-19}$ plus $\log(43/19)=0.8168$, to the four digits printed.

- *Between the classes.* At $\ell=0.8$ both sector matrices have the form $M_0-c_2H_2-c_3H_3-c_2^2H_4$ with $c_2=\chi(2)$ and $c_3=\chi(3)$, so $(c_2,c_3)$ can range over $[-1,1]^2$. We took 81 points with step $1/4$ ($K=40$, 40 digits); at the nine lattice points this reproduces the findings above. In every column $c_2$ the odd ground state has no imaginary zero below a threshold in $c_3$ and one pair above it, and the threshold lies between $0.25$ and $1$. The odd sector is lower only in a corner around $(-1,-1)$, at 10 points, all with $c_2,c_3\le-1/4$. Across every pair boundary on the grid $W(0)$ keeps its sign and $g'(\ell)$ changes sign, so these pairs are born at infinity in the sense of Section 4. A birth at infinity depends on the frequency cutoff, so the thresholds may move with $K$.

Over all 54 cases of Sections 5.4 and 5.5, imaginary pairs occurred only at $\ell\le1.0$ (8 of 33), and never at $\ell\ge1.2$ (0 of 21) [num].

**5.6 The odd sector of $\zeta$.** The same code with $\chi=1$, the archimedean factor of $\zeta$, and the pole term, which enters the odd form as $-2\widehat f(i/2)^2$ [8, §6], reproduces Zhu's sector floors [8, Table 1]. At $\ell=0.6,0.8,1.0,1.2$ ($K=60$, 60 digits) the ratio of the least odd to the least even eigenvalue is 397, 896, 2465 and 5689, against his 371, 949, 2535 and 5997. At $\ell=0.8$ our two floors, $1.68\times10^{-14}$ and $1.87\times10^{-17}$ ($1.67\times10^{-14}$ and $1.74\times10^{-17}$ at $K=100$), lie inside his certified enclosures. So the odd ground state of $\zeta$ is not the global one at these windows, and Theorem 1 says nothing about it, but Theorem 2 applies. At all four windows the odd ground state has no sign change on $(0,\ell)$, so by Corollary 4 every zero of its transform is real. The count of zeros of $W$ agrees: 59 of 59 on $(0,\infty)$, and 99 of 99 at $K=100$, $\ell=0.8$. The same holds at $\ell=1.6$ ($K=100$, 120 digits; odd floor $1.18\times10^{-107}$, even $1.83\times10^{-111}$): no sign change, and 99 of 99 zeros of $W$ on $(0,\infty)$. At $\ell=2.0$ the floors fall below $10^{-120}$, so we used 300 digits (odd floor $2.22\times10^{-158}$, even $1.77\times10^{-161}$). The odd ground state again has no sign change. Here the grid count finds only 91 zeros of $W$ on $(0,\infty)$, because 43 of the 99 lie beyond the last pole $(K\pi/\ell)^2$, some closer together than the grid. The roots of the numerator of $W$, computed directly at 700 digits, are 99 positive reals. The coefficients of $\zeta$ at $n=2,3,4$ are those of the class $\chi(2)=\chi(3)=1$, whose odd ground states have imaginary pairs at $\ell=0.6$ and $0.8$, so the difference comes from the pole term and the archimedean factor.

Osman [10, §2] computed this odd ground state at $L=\log c$ for $c=13,17,19,23$ ($\ell=1.28$ to $1.57$). He found that the sign changes of its transform on $(0.1,50)$, and on $(0.1,100)$ at $c=13$, agree with the ordinates of $\zeta$ in that range, one for one, to within $3\times10^{-15}$. A scan of the real axis cannot see zeros on $i\mathbb R$. At $c=13$ and $c=23$ ($K=100$, 160 digits) the odd ground state has no sign change on $(0,\ell)$, so by Corollary 4 its transform has no zeros off the real axis there either; the count of $W$ agrees (99 of 99). At $c=13$ the ratio of the odd to the even floor is 8829, against 8460 from his values at $N=60$ [10, §2.2].

## 6. Remarks

1. *Why are the zeros real at longer windows?* The ground states need not be positive (Section 5.4). Corollary 4 covers 20 of the 21 cases there; the exception is $\chi_{-20}$, with two sign changes. But it only moves the question to the sign pattern of the ground state, which we cannot explain either. For an odd $f$ the form couples $t$ and $u$ in $(0,\ell)$ through $k(t-u)-k(t+u)$, where $k$ is the kernel of the form. The prime terms of $k$ are point masses $-\chi(n)\Lambda(n)n^{-1/2}$ at $\pm\log n$, and $k(t-u)-k(t+u)$ has no sign; de Andrade Silva [11] shows the corresponding mixed signs of the per-prime terms for the odd vector $\sinh(x/2)$. So the Beurling–Deny criterion and the Perron–Frobenius argument for this kernel, which would make $f$ one-signed, do not apply, and Section 5.4 shows that their conclusion can fail. By Proposition 7 the question is whether the polynomial $R$ has a negative zero. The random model of Section 5.1, where the $b_j$ are independent, shows that the Loewner structure alone does not exclude one, and Section 5.5 shows that Weil forms have one at some short windows. A sign condition on the $b_j$ does exclude it [U]. If $\varphi_j=b_j/j$ is strictly decreasing, every off-diagonal entry of $Q^-=J\Lambda^-J$ is negative, so by Perron–Frobenius the least odd eigenvector $y$ has positive coordinates. Then $R(x)/\prod_i(i^2-x)=\sum_j(Jy)_j/(j^2-x)$ changes sign in each of the $N-1$ intervals between consecutive nodes, so $R$ has no negative zero, whatever the diagonal. When $\varphi$ increases instead, 661 of 800 random least odd vectors have an imaginary pair, against 242 of 800 for independent $b_j$ ($N=3,5,8,12$, 200 draws each, $a_j$ as in Section 5.1) [num]. The Weil forms of Section 5 are far from this case. In the basis of Lemma 6(i), that is, the sine basis after the sign change $(-1)^{j+k}$ of Section 5.3, $\varphi$ is not monotone and 24% to 41% of the off-diagonal entries are positive ($\zeta$ and $\chi_{-4}$ at $\ell=0.8,1.2,1.6$, $\chi_{-8}$ at 1.6; $K=40$). The coordinates of the odd ground state of $\zeta$ change sign 5, 12 and 21 times at $\ell=0.8,1.2,1.6$ (200 digits), although its $f$ has no sign change there [num]. A reason would therefore have to use the values $b_j$ in another way and depend on the window; we found negative zeros only for $\ell\le1.0$. We do not have one.
2. *What this does not give.* Paper 80 [2] measures how closely the real zeros of the ground state's transform track the ordinates of $L(s,\chi)$. Theorems 2 and 3 say nothing about where the real zeros lie. For $\zeta$ and assuming the Riemann hypothesis, Suzuki [7] proves that the spectral entire functions of a self-adjoint extension built from the localized Weil form converge to the canonical product over the ordinates as the window grows; nothing of that kind is proved here.

## References

[1] A. Connes, W. D. van Suijlekom, Quadratic forms, real zeros and echoes of the spectral action, arXiv:2511.23257.
[2] span_of_dilates, An equilibrium conjecture for how far the ground state of a truncated Weil form tracks the zeros, with measurements for Dirichlet L-functions, rStation Paper 80.
[3] J. Agler, Z. A. Lykova, N. J. Young, The boundary Carathéodory–Fejér interpolation problem, J. Math. Anal. Appl. (2011); arXiv:1011.1399.
[4] D. Sarason, Nevanlinna–Pick interpolation with boundary data, Integral Equations Operator Theory 30 (1998).
[5] S. Karlin, Total Positivity, Vol. I, Stanford University Press, 1968.
[6] Y. Shi, Construction of finite Hilbert–Pólya matrices from Weil's explicit formula, arXiv:2609.04908.
[7] M. Suzuki, Spectral convergence associated with Weil's quadratic form, arXiv:2610.32704.
[8] X. Zhu, Weil positivity in compact windows: a finite reduction, certified two-sided bounds, and a Landau–Widom decay law, arXiv:2608.24827v2.
[9] A. Groskin, High-precision approximation of Riemann zeros via the truncated Weil form, arXiv:2605.20224v4.
[10] M. Osman, A probe of the odd parity sector of the CvS Galerkin matrix, numerical report (August 2026), github.com/Osman209/prime-number-studies, file papers/odd_parity_sector.md.
[11] B. W. de Andrade Silva, A Loewner divided-difference formula for the prime contribution in the localized Weil quadratic form, and a parity sign law, Zenodo (2026), doi:10.5281/zenodo.20710075.
