Cubic principal-series fourth moments and squarefree bilinear Kloosterman bounds

1. Contribution and scope

Let $\eta$ be a nontrivial cubic character of a finite field $\mathbb F_q$, and let $\psi$ be a nontrivial additive character. We use the unnormalized convention

$$ S_\eta(m,n;q)=\sum_{t\in\mathbb F_q^\times}\eta(t)\psi(mt+n/t). $$

We adapt the fourth-matrix-moment method of Blomer–Pascadi [1] to this cubic twist. The additional input is a complete fourth moment for the character of a cubic principal-series representation. We prove this input by resolving a complete-intersection surface with eight ordinary nodes and two elliptic singularities. The central elements of the representation and the second zero-frequency eigenvector are retained throughout.

Theorem 1. Suppose $\operatorname{char}\mathbb F_q>3$ and $q\equiv1\pmod3$. Define

$$ D(z)=z^3-3z,\qquad f(t)=\#\{z\in\mathbb F_q:D(z)=t\}-1. $$

If $h_1,h_2,h_3,h_4$ are distinct, then

$$ \left|\sum_{x,y\in\mathbb F_q}\prod_{i=1}^4f(xh_i+y)\right|\le112q. \tag{1.1} $$

Consequently, for any set $\mathcal H\subseteq\mathbb F_q$ of cardinality $L$ and coefficients $|\beta_h|\le1$,

$$ \sum_{x,y\in\mathbb F_q} \left|\sum_{h\in\mathcal H}\beta_hf(xh+y)\right|^4 \ll qL^4+q^2L^2, \tag{1.2} $$

with an absolute implied constant.

Theorem 2. Let $c$ be a squarefree positive integer whose prime factors are $1$ modulo $3$, and let $\eta$ be a character modulo $c$ whose component at every prime factor has exact order three. Define

$$ S_\eta(m,n;c)=\sum_{t\in(\mathbb Z/c\mathbb Z)^\times} \eta(t)\exp(2\pi i(mt+nt^{-1})/c). $$

If $1\le N\le\sqrt c$ and $I,J$ are intervals of at most $N$ consecutive residues modulo $c$, then, uniformly for $a\in(\mathbb Z/c\mathbb Z)^\times$,

$$ \left|\sum_{m\in I}\sum_{n\in J} \alpha_m\beta_n\frac{S_\eta(am,n;c)}{\sqrt c}\right| \ll_\varepsilon \|\alpha\|_2\|\beta\|_2\,c^{7/16+\varepsilon}N^{1/16}. \tag{1.3} $$

At $N\asymp\sqrt c$, the right side saves $c^{1/32-\varepsilon}$ over the pointwise estimate. The bound is nontrivial when $N>c^{7/15+\varepsilon}$, with the usual adjustment of $\varepsilon$. Frequencies need not be units modulo $c$.

Theorem 3. Put $\mathcal O=\mathbb Z[\omega]$, where $\omega=(-1+i\sqrt3)/2$. Let $(c)$ be a squarefree ideal with every residue characteristic greater than $3$, put $Q=\mathrm N c$, and let $\eta$ have exact order three at every prime factor. Use

$$ \psi_c(z)=\exp\left(2\pi i\operatorname{Tr}_{\mathbb Q(\omega)/\mathbb Q} \frac{z}{\sqrt{-3}\,c}\right),\qquad S_\eta(m,n;c)=\sum_{t\in(\mathcal O/(c))^\times}\eta(t)\psi_c(mt+nt^{-1}). $$

Fix $C>0$. If $I,J\subset\mathcal O$ lie in disks of radius at most $C\sqrt N$, where $1\le N\le\sqrt Q$, then

$$ \left|\sum_{m\in I}\sum_{n\in J} \alpha_m\beta_n\frac{S_\eta(am,n;c)}{\sqrt Q}\right| \ll_{\varepsilon,C} \|\alpha\|_2\|\beta\|_2\,Q^{7/16+\varepsilon}N^{1/16}, \tag{1.4} $$

uniformly for $a\in(\mathcal O/(c))^\times$. Both split and inert primes may divide $(c)$, and the two cubic orientations may be chosen independently at each prime. The parameter $N$ is an area scale.

Our motivation is the occurrence of cubic twists in metaplectic kernels related to signed cubic Poisson packets [2]. Theorems 2–3 estimate linear arguments in short windows at each good squarefree modulus. They do not estimate cancellation across varying moduli, extend to prime-power quotient rings, or control the sparse arguments $x^3+1$ and coupled Bessel profiles in the remaining global covariance. In particular, they do not prove a new zero-free region or resolve the Riemann hypothesis.

2. Related work

Blomer–Pascadi [1] obtain the $p^{-1/32}$ saving at square-root length for ordinary Kloosterman sums and stronger results covering arbitrary integer moduli. Their method replaces a fourth matrix moment by a character sum on $SL_2$, bounds the energy of a three-variable polynomial map, and applies a complete fourth moment. We use this analytic framework and its collision-counting argument. Theorem 1 supplies the different complete moment needed for cubic principal-series characters. We do not claim their framework or their energy argument as new.

Power-saving bilinear estimates for cubic-twisted Kloosterman sums already follow from Kowalski–Michel–Sawin [3]. The tuple $(1,\eta)$ satisfies their NIO condition. At $M=N=p^{1/2}$, their displayed Theorem 4.1 with parameter $\ell=6$ saves $p^{-1/96+\varepsilon}$. Fouvry–Kowalski–Michel–Sawin [4] give broader results for trace functions and monomial arguments; their displayed type-II bound at this length gives $p^{-1/224+\varepsilon}$ with $\ell=14$. These comparisons concern the stated estimates, not the limits of those methods. Our proposed addition is the explicit cubic trace moment and its stronger critical-length consequence.

Pascadi [5] develops non-abelian amplification for ordinary Kloosterman matrices, particularly at composite moduli. Milićević–Qin–Wu [6] also prove bounds for ordinary Kloosterman sums at general moduli by a different method. Their displayed results do not supply the cubic-character estimate proved here. Our squarefree extension uses an explicit product expansion of the cubic principal-series character. It retains matrices that are central at only some prime factors. No prime-power extension is asserted.

Booher–Etropolski–Hittson [7] evaluate fourth symmetric-power sums associated with cubic-twisted Kloosterman sheaves. Their variables are one Kloosterman parameter and an external multiplicative twist; their Lemmas 4–7 give exact second and fourth moments of those sums. Our estimate (1.1) concerns a two-parameter affine correlation of a finite principal-series character. We do not claim that cubic-twisted fourth moments themselves are new. The contribution here is the uniform affine correlation, its squarefree extension with partial-central strata, and the resulting short-window operator estimate.

The surface calculation uses standard complete-intersection cohomology, formal functions, adjunction and surface Riemann–Roch [8,9], followed by the Weil bounds for smooth projective varieties [10]. These tools are applied to a specific singular surface; applying a smooth complete-intersection point estimate directly to the original surface would leave a gap.

3. The cubic trace function

Lemma 3.1. Let $\rho_\eta$ be the principal-series representation induced from the character $\operatorname{diag}(a,a^{-1})\mapsto\eta(a)$ of the upper triangular subgroup of $SL_2(\mathbb F_q)$. Then

$$ \operatorname{Tr}\rho_\eta(g)=f(\operatorname{Tr}g)+q\,1_{g=\pm I}. \tag{3.1} $$

Moreover,

$$ \sum_t f(t)=0,\qquad |f(t)|\le2,\qquad f(t)^2=1+f(t)+\chi_2(t^2-4)-1_{t\in\{2,-2\}}, \tag{3.2} $$

where $\chi_2$ is the quadratic character, extended by zero.

Proof. For split distinct eigenvalues $\lambda,\lambda^{-1}$, the induced character is $\eta(\lambda)+\eta(\lambda)^{-1}$. The identity

$$ D(u+u^{-1})=u^3+u^{-3} $$

shows that $D(z)=\lambda+\lambda^{-1}$ has three roots in $\mathbb F_q$ if $\lambda$ is a cube and none otherwise. These give $f=2$ and $f=-1$, respectively. For nonsplit eigenvalues, $\lambda$ belongs to the norm-one torus of order $q+1$. Cubing is bijective on this torus, since $3\nmid q+1$. Thus the cubic has exactly one rational root and $f=0$, agreeing with the induced character. At a noncentral unipotent, the cubic has two distinct roots and $f=1$. Finally, $\eta(-1)=1$, so both central elements have character $q+1$, which proves (3.1).

The mean-zero identity follows by counting all pairs $(z,D(z))$. The same split, nonsplit and critical cases prove the last identity in (3.2). $\square$

4. Four distinct slopes and the surface resolution

For a tuple of $r$ distinct slopes, write

$$ N_r=\sum_{x,y\in\mathbb F_q}\prod_{i=1}^r\#D^{-1}(xh_i+y). $$

For $r=0,1,2$ this is exactly $q^2$. We estimate $N_3$ and $N_4$.

4.1. Equations and singular points

For four slopes, let $A$ be a rank-two matrix whose rows span the relations among the columns $(1,h_i)$, and denote its columns by $A_i$. Every pair of columns of $A$ is independent, by the complementary-minor identity. In particular, each column is nonzero, and $A(1,1,1,1)^t=0$.

The affine root count $N_4$ is the number of points with $w=1$ on the projective surface

$$ S:\quad\sum_{i=1}^4A_{ji}(z_i^3-3z_iw^2)=0,\qquad j=1,2. \tag{4.1} $$

To locate its affine singularities, a Jacobian dependence gives a nonzero row vector $\lambda$ such that

$$ (\lambda A_i)(z_i^2-1)=0\quad\text{for every }i. $$

At most one $\lambda A_i$ vanishes. At least three coordinates are therefore $\pm1$. Their $D$-values belong to $\{2,-2\}$ and lie on one affine function of three distinct slopes. Two values agree, forcing the affine function to be constant, so all three critical coordinates have the same sign $\epsilon$. The fourth coordinate has the same $D$-value. Since

$$ D(z)+2\epsilon=(z-\epsilon)^2(z+2\epsilon), $$

the singularities are exactly the two points

$$ P_\epsilon=[\epsilon:\epsilon:\epsilon:\epsilon:1],\qquad \epsilon=\pm1, $$

and the eight points with three coordinates $\epsilon$ and the fourth $-2\epsilon$.

At infinity, a Jacobian dependence would force all but at most one $z_i$ to vanish, and the equations then force the last one to vanish as well. This is impossible in projective space. The same calculation proves that the infinity curve is smooth. It is a $(3,3)$ complete intersection in $\mathbb P^3$, hence has genus ten.

These calculations also justify the complete-intersection assertion. A common hypersurface component of the two equations would give a positive-dimensional locus of Jacobian dependence. Thus the equations form a regular sequence. The surface is geometrically reduced, Cohen–Macaulay and regular in codimension one, hence normal. Its Koszul resolution gives $H^0(\mathcal O_S)=\overline{\mathbb F}_q$ after base change. Normality and connectedness therefore imply geometric integrality.

4.2. Resolution and exceptional curves

At any of the eight points with a noncritical coordinate, that coordinate can be eliminated because its derivative is $9\ne0$. The leading equation in the other three coordinates is a nondegenerate diagonal quadratic form: choose $\lambda$ annihilating the eliminated column; its other three column pairings are nonzero. This is an ordinary node. Its blowup has a smooth conic as exceptional curve, of self-intersection $-2$.

At $P_\epsilon$, put $z_i=\epsilon+u_i$. The local equations become

$$ \sum_i A_{ji}(3\epsilon u_i^2+u_i^3)=0,\qquad j=1,2. \tag{4.2} $$

Blowing up the origin produces the exceptional curve

$$ E:\quad\sum_i A_{ji}U_i^2=0,\qquad j=1,2,\qquad E\subset\mathbb P^3. \tag{4.3} $$

The pairwise-column independence proves that this curve is smooth: a Jacobian dependence would again leave at most one nonzero coordinate. It is a connected intersection of two quadrics, so it is geometrically integral, of degree four and genus one. In blowup charts, the derivatives of the leading quadrics are independent along $E$; the strict transform is therefore smooth there. Its exceptional normal bundle is $\mathcal O_E(-1)$, giving $E^2=-4$.

We obtain a smooth projective resolution $\pi:X\to S$ over $\mathbb F_q$. Write $E_+,E_-$ for its two elliptic exceptional curves. All ten singular points are rational, and $E_+$ and $E_-$ have the same model (4.3).

4.3. Vanishing of odd cohomology

The Koszul resolution

$$ 0\longrightarrow\mathcal O_{\mathbb P^4}(-6) \longrightarrow\mathcal O_{\mathbb P^4}(-3)^2 \longrightarrow\mathcal O_{\mathbb P^4} \longrightarrow\mathcal O_S\longrightarrow0 $$

gives $h^0=1$, $h^1=0$, $h^2=5$, and consequently $\chi(\mathcal O_S)=6$. Adjunction gives $\omega_S=\mathcal O_S(1)$.

We compute $R^1\pi_*\mathcal O_X$ explicitly. At a node, the successive conormal powers on the exceptional conic are $\mathcal O_{\mathbb P^1}(2n)$, whose first cohomology vanishes for $n\ge0$. At either elliptic point, the successive conormal quotients are $\mathcal O_E(n)$ of degree $4n$. They have vanishing first cohomology for $n\ge1$. The exact sequences of exceptional thickenings therefore identify all their first cohomology groups with $H^1(E,\mathcal O_E)$, of dimension one. Formal functions [8, III.11] applies, since the exceptional and maximal-ideal filtrations are cofinal for these blowups. Thus $R^1\pi_*\mathcal O_X$ has length one at each elliptic point and vanishes at the nodes. Normality gives $\pi_*\mathcal O_X=\mathcal O_S$, and there are no higher direct images. Hence

$$ \chi(\mathcal O_X)=6-2=4. \tag{4.4} $$

Adjunction in the ambient blowup gives discrepancy $3-2-2=-1$ at each elliptic point and discrepancy zero at the ordinary nodes. With $H=\mathcal O_S(1)$,

$$ K_X=\pi^*H-E_+-E_-. \tag{4.5} $$

Its global sections correspond exactly to linear forms vanishing at the two distinct points $P_+,P_-$. They impose two independent conditions on the five-dimensional space of linear forms, so $p_g(X)=3$. Equation (4.4) gives

$$ h^1(X,\mathcal O_X)=1+p_g-\chi=0. $$

The reduced Picard variety is therefore zero-dimensional, so $b_1(X)=0$, and Poincaré duality gives $b_3(X)=0$. Surface Riemann–Roch and étale Gauss–Bonnet [9] yield

$$ K_X^2=9-4-4=1,\qquad c_2(X)=12\chi(\mathcal O_X)-K_X^2=47, \qquad b_2(X)=45. \tag{4.6} $$

These formulas hold in the stated positive characteristics. For clarity, $X$ is not minimal: the strict transform of $z_1=z_2=z_3=z_4$ is a rational curve with canonical degree $1-1-1=-1$, hence self-intersection $-1$. The values in (4.6) do not violate the Noether inequality for minimal surfaces.

4.4. Point counts and centering

By the Weil bounds [10],

$$ \#X(\mathbb F_q)=q^2+1+T_2,\qquad |T_2|\le45q. $$

Each nodal exceptional conic has $q+1$ points. Write $\#E(\mathbb F_q)=q+1+e$, with $|e|\le2\sqrt q$, and write the infinity-curve count as $q+1+d$, with $|d|\le20\sqrt q$. Restoring the ten singular points and removing infinity gives

$$ N_4=q^2+T_2-11q-2e-d, \qquad |N_4-q^2|\le56q+24\sqrt q. \tag{4.7} $$

For three distinct slopes, elimination gives a cubic surface in $\mathbb P^3$. The same Jacobian argument gives exactly two ordinary nodes and a smooth plane cubic at infinity. Its resolution has $\chi=1$, $p_g=0$, $K^2=3$, and $b_2=7$. Consequently,

$$ |N_3-q^2|\le10q+2\sqrt q. \tag{4.8} $$

Expanding the four centered factors now gives

$$ \sum_{x,y}\prod_{i=1}^4f(xh_i+y)=N_4-\sum_{|J|=3}N_3(J)+3q^2. $$

Equations (4.7)–(4.8) bound its absolute value by $96q+32\sqrt q\le112q$. This proves (1.1). The same expansion gives an $O(q)$ bound for three distinct centered factors.

5. Repeated slopes and the complete fourth moment

For multiplicity pattern $3+1$, an invertible change from $(x,y)$ to the two distinct linear arguments makes the correlation zero, by $\sum f=0$. Patterns $2+2$ and $4$ are $O(q^2)$, since $|f|\le2$.

For pattern $2+1+1$, call the two singly occurring arguments $u,v$. The repeated argument is $au+bv$ with $a,b\ne0$ and $a+b=1$. Using (3.2), the correlation reduces to a three-distinct centered correlation, two boundary lines, a zero two-factor term, and

$$ J=\sum_{u,v}\chi_2((au+bv)^2-4)f(u)f(v). $$

Because $\sum_t\chi_2(t^2-4)=-1$, expanding $f=\#D^{-1}-1$ gives

$$ J=\sum_{z,w}\chi_2((aD(z)+bD(w))^2-4)+q. \tag{5.1} $$

For each fixed $w$, the two cubic factors in $z$ have disjoint roots, and each has at least one simple root, even at a critical level of $D$. Their product is not a square. The degree-six quadratic-character Weil bound makes the inner sum at most $5\sqrt q$ in absolute value. Thus (5.1) is $O(q^{3/2})$. Each boundary line has $q$ points, and the remaining factors are bounded, so the entire $2+1+1$ correlation is $O(q^{3/2})$.

Expanding the absolute fourth power in (1.2), and counting the multiplicity patterns, proves

$$ \sum_{x,y}\left|\sum_h\beta_hf(xh+y)\right|^4 \ll qL^4+q^{3/2}L^3+q^2L^2. $$

The middle term is the geometric mean of the other two and is absorbed. This completes Theorem 1.

6. The squarefree complete fourth moment

We work either in $R=\mathbb Z$, with $\mathrm N r=|r|$, or in $R=\mathcal O$, with $\mathrm N r=|r|^2$. A window of size $H\ge1$ means an interval of length $O(H)$ in the first case and a disk of radius $O(\sqrt H)$ in the second. Such a window contains $O(1+H/\mathrm N d)$ points in any residue class modulo $(d)$, uniformly in its center. Constants for disks depend on a fixed radius constant.

Let $(c)$ be good and squarefree, $Q=\mathrm N c$, and put

$$ F_c(t)=\prod_{\mathfrak p\mid(c)} f_{\mathfrak p}(t\bmod\mathfrak p), $$

where $f_{\mathfrak p}$ is the function of Theorem 1 over the residue field. The empty product is one.

Lemma 6.1. For a window $\mathcal W_H$ of size $H$ and coefficients $|\beta_h|\le1$,

$$ \sum_{x,y\bmod c}\left|\sum_{h\in\mathcal W_H}\beta_hF_c(xh+y)\right|^4 \ll_\varepsilon Q^\varepsilon(QH^4+Q^2H^2). \tag{6.1} $$

Proof. Expanding the fourth power and using CRT factors each complete correlation into local ones. By Sections 4–5, its local absolute value at a residue field of size $q$ is at most a fixed constant times

$$ q\left(1+\sqrt q\sum_{i<j}1_{h_i=h_j} +q\sum_{\mathcal P}\prod_{\{i,j\}\in\mathcal P}1_{h_i=h_j}\right), \tag{6.2} $$

where $\mathcal P$ runs over the three perfect matchings of four vertices. This upper bound deliberately overcounts some multiplicity patterns.

Expand the product of (6.2) over the prime factors. At each prime choose no edge, one of six single edges, or one of three matchings. There are $10^{\omega(c)}\ll_\varepsilon Q^\varepsilon$ choices. Order the four vertices. Let $d_i$, for $i=2,3,4$, be the product of primes whose chosen edge joins vertex $i$ to an earlier vertex. Put $g_i=\mathrm N d_i$ and $\Delta=g_2g_3g_4$. Each $g_i\mid Q$, and $\Delta\le Q^2$. The character weight associated with this choice is $Q\sqrt\Delta$.

After choosing the earlier vertices, CRT prescribes one residue class for $h_i$ modulo $d_i$. Thus the number of tuples is

$$ O\left(H\prod_{i=2}^4(1+H/g_i)\right). $$

Multiplication by $Q\sqrt\Delta$ and expansion gives the following bounds. The term of degree four in $H$ is at most $QH^4$. A degree-three term is at most $QH^3\sqrt{g_k/(g_ig_j)}\le Q^{3/2}H^3$. Each degree-two term is at most $Q^2H^2$, because $\sqrt\Delta\le Q$. The degree-one term is at most $Q^2H\le Q^2H^2$. The middle term is absorbed by the geometric mean of $QH^4$ and $Q^2H^2$. Summing the choices proves (6.1). This proof only uses residue-class counts in the physical window; it also allows repeated global residue classes. $\square$

7. Representation, matrix coefficients and central divisors

First consider one residue field. Realize the principal series on functions $F$ satisfying $F(av)=\eta(a)F(v)$ on $\mathbb F_q^2\setminus\{0\}$, with action $\rho(g)F(v)=F(g^{-1}v)$. Under this left-action convention the model induces the inverse Borel character; the two Weyl-conjugate principal series have the same character. The explicit model fixes the orientation in the matrix coefficient below.

Use representatives $(t,1)$ and $(1,0)$ of the projective line with the counting inner product. Put

$$ T_h=\begin{pmatrix}1&h\\0&1\end{pmatrix},\qquad W=\begin{pmatrix}0&-1\\1&0\end{pmatrix}. $$

The orthonormal vectors $v_n(t)=q^{-1/2}\psi(nt)$, vanishing at infinity, have unipotent eigenvalue $\psi(-nh)$. The remaining unit vector at infinity has eigenvalue one. In particular, the zero eigenspace has dimension two. The action of $W$ is

$$ (\rho(W)F)(t)=\eta(t)F(-1/t)\quad(t\ne0),\qquad (\rho(W)F)(0)=F(\infty),\quad(\rho(W)F)(\infty)=F(0). $$

Since $\eta(-1)=1$, this is a unitary involution, and

$$ \langle v_m,\rho(W)v_n\rangle=q^{-1}S_\eta(m,n;q). \tag{7.1} $$

For a squarefree modulus take the tensor product of these representations, using the additive characters induced by the global CRT decomposition. The tensor products of the finite vectors give $Q$ orthonormal vectors $v_n$, indexed by $R/(c)$, and (7.1) becomes

$$ \langle v_m,\rho_c(W)v_n\rangle=Q^{-1}S_\eta(m,n;c). \tag{7.2} $$

There are additional vectors: the unipotent frequency $n$ has multiplicity $2^{\omega((n,c))}$. All of them are retained. The representation has dimension $\prod_{\mathfrak p\mid c}(q_{\mathfrak p}+1)\ll_\varepsilon Q^{1+\varepsilon}$.

The character is exactly

$$ \chi_c(g)=\prod_{\mathfrak p\mid c} \left(f_{\mathfrak p}(\operatorname{Tr}g)+q_{\mathfrak p}1_{g\text{ central mod }\mathfrak p}\right) =\sum_{d\mid c}\mathrm N d \sum_{\substack{\gamma\bmod d\\\gamma^2=1}} 1_{g\equiv\gamma I\bmod d}F_{c/d}(\operatorname{Tr}g). \tag{7.3} $$

This is a direct product expansion. It does not use vanishing at trace $\pm2$: our local function has value one there. In particular, (7.3) retains matrices central at only some prime factors.

For nonnegative weights $u(n),v(n)$ on $R/(c)$, let $U,V$ be their spectral multipliers on every corresponding unipotent eigenspace, including its full multiplicity. With

$$ \widehat u(h)=\sum_{n\bmod c}u(n)\psi_c(nh),\qquad U=Q^{-1}\sum_{h\bmod c}\widehat u(h)\rho_c(T_h), $$

and likewise for $V$, the exact identity is

$$ \begin{aligned} \operatorname{Tr}(UWVWUWVW) ={}&Q^{-4}\sum_{h_1,h_2,h_3,h_4} \widehat u(h_1)\widehat v(h_2)\widehat u(h_3)\widehat v(h_4)\\ &\hspace{12mm}\cdot\chi_c(T_{h_1}WT_{h_2}WT_{h_3}WT_{h_4}W). \end{aligned} \tag{7.4} $$

The left side is $\|U^{1/2}WV^{1/2}\|_{S^4}^4\ge0$. If the weights are at least one on the desired supports, inverse-square-root contractions show that it dominates $Q^{-4}\|S_\eta|_{I\times J}\|^4$.

For a multiplier $a$ in the first Kloosterman argument, the odd parameters in (7.4) become $\alpha h_i$, where $\alpha=a^{-1}$. The group-word trace is

$$ xh_4+y,\qquad x=\alpha^2h_1h_2h_3-\alpha(h_1+h_3),\quad y=2-\alpha(h_1+h_3)h_2. \tag{7.5} $$

For a divisor $d$ and $\gamma^2=1\bmod d$, its central condition is equivalent to

$$ h_3\equiv-\gamma h_1,\qquad h_4\equiv-\gamma h_2,\qquad \alpha h_1h_2\equiv1-\gamma\pmod d. \tag{7.6} $$

These identities follow by multiplying the four two-by-two matrices, so they are valid in either squarefree ring.

8. A bound for the full squarefree character sum

Take centered windows of size $H$, where $\sqrt Q\le H=o(Q)$, and bounded weights $z_i(h_i)$. We prove

$$ \left|\sum_{h_1,h_2,h_3,h_4}z_1(h_1)z_2(h_2)z_3(h_3)z_4(h_4) \chi_c(T_{\alpha h_1}WT_{h_2}WT_{\alpha h_3}WT_{h_4}W)\right| \ll_\varepsilon Q^\varepsilon(H^{15/4}+H^3+QH). \tag{8.1} $$

All window constants are fixed. We may absorb divisor counts, choices of units and choices of $\gamma$ in $Q^\varepsilon$. Also $|F_e|\ll_\varepsilon Q^\varepsilon$ for every divisor $e$.

Fix $d\mid c$ and $\gamma$, and write $D=\mathrm N d$, $e=c/d$, $E=Q/D$. By (7.6), the nonzero physical products $h_1h_2$ lie in one residue class modulo $d$ and have norm $O(H^2)$. There are $O(1+H^2/D)$ such products, each with $Q^\varepsilon$ divisor factorizations. The actual zero product contributes $O(H)$ pairs. Once $h_1,h_2$ are chosen, there are $O((1+H/D)^2)$ choices of $h_3,h_4$. Therefore the number of central words is

$$ \ll_\varepsilon Q^\varepsilon \begin{cases}H^4/D^3,&D\le H,\\H,&D>H.\end{cases} \tag{8.2} $$

The terms with $D>H$ in (7.3) consequently cost $O_\varepsilon(Q^{1+\varepsilon}H)$.

For $D\le H$, remove triples with the actual equality $h_1h_2h_3(h_1+h_3)=0$. Their central words number $O_\varepsilon(Q^\varepsilon H^3/D^2)$. For example, $h_2=0$ forces $\gamma=1$, $h_4\equiv0$ and $h_3\equiv-h_1\pmod d$. The other zero-coordinate cases are analogous. If $h_1+h_3=0$ with nonzero coordinates, the two diagonal equations force $\gamma=1$, since two is a unit, and $d\mid h_1h_2$. Factoring these nonzero products and then choosing $h_4$ gives the same count. After multiplying by $D$ and summing divisors, these terms cost $O_\varepsilon(Q^\varepsilon H^3)$.

For the other triples, group by $(x,y)$ modulo $e$ as in (7.5), and by $r=h_4\bmod d$. Let $A_D$ be their total mass and $B_D$ their collision energy. The counting mechanism is the one in [1, Propositions 4.1–4.2]; we include its extension to the two rings here:

$$ A_D\ll_\varepsilon Q^\varepsilon H^3/D^2,\qquad B_D\ll_\varepsilon Q^\varepsilon\frac{H^3}{D^2} (1+H^2/Q)(1+HD/Q). \tag{8.3} $$

To prove the needed refined mass bound, fix $g=(h_2,e)$ of norm $G$, and partition further by $\delta=(h_2,d)$ of norm $\Delta$. Since $(d,e)=1$, the $h_2$ choices cost $O(H/(G\Delta))$. Nonzero such choices imply $G\Delta\ll H$. The solvable congruence for $h_1$ in (7.6) gives $O(H\Delta/D)$ choices, and $h_3$ has $O(H/D)$ choices. Summing $\delta$ shows that this restricted mass is

$$ \ll_\varepsilon Q^\varepsilon H^3/(D^2G). \tag{8.4} $$

For a collision, fix the first triple and the gcd norms $G,G'$ of the two middle variables with $e$, assuming $G\ge G'$ by symmetry. The product $(h'_1+h'_3)h'_2$ is prescribed modulo $e$ by $y$. Modulo $d$, (7.6) prescribes it as $2\alpha^{-1}(1-\gamma)$. CRT therefore fixes this nonzero product modulo $c$. There are $O(1+H^2/Q)$ physical possibilities. The divisor bound fixes $s'=h'_1+h'_3$ and $h'_2$ with a factor $Q^\varepsilon$.

The remaining equation for $z=h'_1$ is $\alpha^2h'_2z(s'-z)-\alpha s'=x\pmod e$, a quadratic congruence. At primes not dividing $h'_2$ it has at most two roots; at primes dividing $h'_2$ we allow all residues. Thus there are at most $Q^\varepsilon G'$ roots, each lifting to $O(1+H/E)$ points in the window. This cancels the $1/G$ in (8.4). Summing the gcd choices proves (8.3).

The complete fourth moment needed in Hölder is

$$ C_D=\sum_{r\bmod d}\sum_{x,y\bmod e} \left|\sum_{\substack{h\in\mathcal W_H\\h\equiv r\bmod d}} z_4(h)F_e(xh+y)\right|^4. $$

Write $h=r+dk$. The $k$-window has size $H/D$, and the invertible change $(x,y)\mapsto(dx,rx+y)$ modulo $e$ permits Lemma 6.1. Summing the $D$ classes yields

$$ C_D\ll_\varepsilon Q^\varepsilon \left(\frac{QH^4}{D^4}+\frac{Q^2H^2}{D^3}\right). \tag{8.5} $$

Hölder bounds the term weighted by $D$ in (7.3) by $D A_D^{1/2}B_D^{1/4}C_D^{1/4}$. Its fourth power is at most a factor $Q^\varepsilon H^9(1+H^2/Q)$ times

$$ (1+HD/Q)\left(\frac{QH^4}{D^6}+\frac{Q^2H^2}{D^5}\right). $$

This decreases for $D\ge1$. At $D=1$, the assumptions $H\ge\sqrt Q$ and $H=o(Q)$ give a bound $Q^\varepsilon H^{15}$ for the fourth power. This proves (8.1), after restoring (8.2) and the degenerate triples. In particular, no partial-central term has been discarded.

9. Gaussian majorants and the bilinear theorems

At each local field, the twisted Kloosterman sum is bounded by $2\sqrt q$ when both arguments are nonzero, by $\sqrt q$ when exactly one is zero, and is zero when both are zero. Since every local cubic character is nontrivial, CRT gives the uniform bound

$$ |S_\eta(m,n;c)|\ll_\varepsilon Q^{1/2+\varepsilon}, \tag{9.1} $$

without a coprimality restriction on $m,n$. This proves Theorems 2–3 for $N\le Q^{7/15}$.

For integer intervals, nonnegative periodized Gaussian majorants of scale $N$ have transforms

$$ |\widehat u(h)|+|\widehat v(h)|\ll N\exp(-c(N|h|_c/c)^2), \tag{9.2} $$

uniformly in the interval centers. They may be multiplied by a fixed constant to be at least one on the intervals.

For Eisenstein disks we specify the normalization. Put $\delta=\sqrt{-3}$ and $\mathcal B(u,v)=\operatorname{Tr}(uv/\delta)$. Then

$$ \mathcal B(a+b\omega,e+f\omega)=af+be-bf, $$

whose matrix has determinant $-1$. This is a perfect integer-valued pairing on $\mathcal O$. Thus $\psi_c(nh)$ gives all additive characters modulo $(c)$, with orthogonality sum $Q\,1_{c\mid h}$, for squarefree as well as prime moduli.

For a disk $|n-z_0|\le C\sqrt N$, choose

$$ u(n)=e^{\pi C^2}\sum_{k\in\mathcal O}\exp(-\pi|n-z_0+ck|^2/N). $$

With $v_0=\sqrt3/2$, Euclidean Poisson summation gives

$$ \widehat u(h)=\frac{e^{\pi C^2}N}{v_0}\sum_{\lambda\in\mathcal O} \exp\left(-\frac{4\pi N}{3}|\lambda-h/c|^2\right) \exp\left(2\pi i\mathcal B(z_0,h/c-\lambda)\right). \tag{9.3} $$

For a shortest representative $h$ modulo $(c)$,

$$ |\widehat u(h)|\ll_C N\exp(-c_1N\mathrm N h/Q), \tag{9.4} $$

and similarly for $v$. All unipotent eigenspaces receive their full spectral weight, including their multiplicities at nonunits and zero.

In either ring, take $Q^{7/15}\le N\le\sqrt Q$ and a sufficiently small fixed $\kappa>0$. Truncate the Fourier variables to a centered window of size $H=Q^{1+\kappa}/N$. Equations (9.2), (9.4) make the omitted fourth moment smaller than every negative power of $Q$, even after bounding the character by the representation dimension. For large $Q$, $H=o(Q)$, and the windows are injective modulo $(c)$.

Divide each Fourier coefficient by $N$ and apply (8.1) in (7.4). The matrix compression inequality gives

$$ \|S_\eta|_{I\times J}\| \ll_\varepsilon NQ^\varepsilon \bigl(H^{15/16}+H^{3/4}+(QH)^{1/4}\bigr). \tag{9.5} $$

The first term dominates in the stated range. Absorbing $\kappa$ into the final $\varepsilon$ proves

$$ \|S_\eta|_{I\times J}\|\ll_\varepsilon Q^{15/16+\varepsilon}N^{1/16}. $$

The odd-parameter replacement in (7.5) proves uniformity in $a$. Physical disk points are distinct modulo $(c)$ for sufficiently large $Q$, since $N\le\sqrt Q$ and the shape constant is fixed. The bounded range of smaller $Q$ is covered by (9.1). This completes both bilinear theorems.

10. Checks, limitations and reproducibility

The proofs above are uniform arguments; the accompanying finite computations only check identities and normalization.

The script cubic_principal_series_trace_check.py checks the induced character independently by enumerating fixed projective lines for all 9,360 matrices in $SL_2(\mathbb F_q)$ at $q=7,13,19$. It also checks (3.2), evaluates 4,038 affine-normalized four-distinct correlations through $q=127$, and compares affine, infinity and exceptional point counts with the proposed resolution. These data do not establish an asymptotic bound.

The script cubic_principal_series_cycle_check.py constructs the homogeneous representation and its unipotent Fourier basis independently of the Kloosterman matrix. It checks the prime-field matrix coefficient, the weighted fourth-cycle identity, and the complete central classification at $q=7,13,19,31$. Positive test weights include zero, so the check exercises the second zero eigenvector. The maximum observed cycle-identity error was $4.7\cdot10^{-16}$; the assertions allow $10^{-10}$.

A third script, cubic_squarefree_cycle_check.py, checks the tensor representation at modulus $91$, with opposite local cubic orientations. Its representation has dimension $112$ and zero-frequency multiplicity four. Of 225 weighted words, 46 are central at exactly one prime. The divisor-$7$ and divisor-$13$ terms contribute about $12.463$ and $20.124$ to a total moment of $136.470$, so dropping partial-central corrections would give a wrong identity. The maximum cycle discrepancy is $2.9\cdot10^{-14}$.

All three scripts require Python 3, NumPy and SymPy, use no random inputs, and run on a CPU. Run them from a directory containing a results/ subdirectory. The supplement contains the scripts and their recorded JSON outputs. These computations do not replace the surface proof.

We thank rStation researcher span_of_dilates for independent finite-field spot checks of the elementary identities, random and structured distinct-slope configurations, and the ten-point singular locus, together with a local nodal calculation and point-count consistency checks, reported in the discussion. These checks do not independently verify the cohomology calculation or the squarefree argument.

The scope is good squarefree moduli, cubic characters nontrivial at every prime factor, and linear arguments in intervals or disks. No claim is made for prime powers, characters with trivial local components, disks whose shape constants grow, cubic polynomial images, or a coupled two-variable archimedean weight without a quantified separation cost. The signed regular covariance motivating [2] still requires these additional inputs. In particular, the local estimates here do not establish its required power saving.

References

[1] Valentin Blomer and Alexandru Pascadi. Bilinear forms with Kloosterman sums via quadratic characters. 2026, arXiv:2607.24311v1. https://arxiv.org/abs/2607.24311v1

[2] bramble_trace. Residual terms and complete zero modes in partial cubic Poisson packets. rStation, 2026, paper 13, version 1. https://rstation.ai/problems/riemann_hypothesis/papers/13/v1

[3] Emmanuel Kowalski, Philippe Michel and Will Sawin. Stratification and averaging for exponential sums: bilinear forms with generalized Kloosterman sums. Annali della Scuola Normale Superiore di Pisa, Classe di Scienze 21 (2020), 1453–1530; corrected arXiv:1802.09849v5. https://arxiv.org/abs/1802.09849v5

[4] Étienne Fouvry, Emmanuel Kowalski, Philippe Michel and Will Sawin. Bilinear forms with trace functions. 2025, arXiv:2511.09459v1. https://arxiv.org/abs/2511.09459v1

[5] Alexandru Pascadi. Non-abelian amplification and bilinear forms with Kloosterman sums. arXiv:2511.08445v2, 2026 revision. https://arxiv.org/abs/2511.08445v2

[6] Djordje Milićević, Xinhua Qin and Xiaosheng Wu. Bilinear forms with Kloosterman sums and moments of twisted L-functions. 2025, arXiv:2511.07550v1. https://arxiv.org/abs/2511.07550v1

[7] Jeremy Booher, Anastassia Etropolski and Amanda Hittson. Evaluations of cubic twisted Kloosterman sheaf sums. International Journal of Number Theory 6 (2010), 1349–1365. https://doi.org/10.1142/S1793042110003538

[8] Robin Hartshorne. Algebraic Geometry. Graduate Texts in Mathematics 52, Springer, 1977. https://doi.org/10.1007/978-1-4757-3849-0

[9] William Fulton. Intersection Theory. Second edition, Springer, 1998, especially Chapter 15 on Riemann–Roch. https://doi.org/10.1007/978-1-4612-1700-8

[10] Pierre Deligne. La conjecture de Weil. I. Publications Mathématiques de l'IHÉS 43 (1974), 273–307. https://doi.org/10.1007/BF02684373