# Fourth traces and mixed moments of Gaussian quartic characters

## 1. Results and previous work

The quartic large sieve has extra terms even when the row and column
lengths agree. For a prescribed convolution coefficient, an operator
norm need not be the most efficient way to use that sieve. We estimate
a fourth trace of the untwisted character matrix and compare it with
the positive density matrix generated by quadratic modulation. The
pointwise size of the original coefficient is essential to that comparison.

Write $q_n=Nn$ for the norm of an ideal of $\mathbb Z[i]$. Odd ideals
are represented by their primary generators. Put
$\chi_n(h)=(h/n)_4$, retaining its zero whenever $(n,h)>1$. A star
means squarefree. Physical rows are ordinary unless specified otherwise.
Fix nonnegative smooth profiles compactly supported in $(0,\infty)$.
For squarefree columns $q_n\asymp B$, let

$$
 A(Y,B)_{h,n}=\Phi(q_h/Y)^{1/2}\chi_n(h),\qquad
 T_4(Y,B)=\operatorname{Tr}\bigl((A(Y,B)^*A(Y,B))^4\bigr).
 \tag{1}
$$

Let $V_4$ denote the same trace when the columns are fourth-power-free.
Bounded column weights and arbitrary fixed column restrictions are
allowed in these definitions. Constants depend on finitely many profile
seminorms, fixed local data and the stated subsidiary exponent.

**Theorem 1.** For $1\le B\le Y$,

$$
 \begin{aligned}
 V_4(Y,B)&\ll_\epsilon(YB)^\epsilon
                 \left(Y^4B+Y^{-2}B^{15/2}\right),\\
 T_4(Y,B)&\ll_\epsilon(YB)^\epsilon
                 \left(Y^4B+Y^{-7/2}B^9\right).
 \end{aligned}
 \tag{2}
$$

If $K_L$ is the Gram with ordinary rows of length $L$ and the full
squarefree column ball $q_n\le CL$, then

$$
                 \operatorname{Tr}K_L^4\ll L^{26/5+\epsilon}.
 \tag{3}
$$

The large-sieve input is de Faveri's fixed-order theorem and coupled
root-number formula [1]. The second trace comes from the inverse-pair
bound [2, Proposition 4]. The matrix organization follows the cubic
trace argument [3]. The new arithmetic step retains the quartic sharp
part inside its radical, and counts its labels by their actual number.
Prime dilation then transfers a long-row trace to equal lengths.

**Theorem 2.** Suppose $c_n$ is supported on all ordinary odd columns
$q_n\le CL$ and
$|c_n|\le M L^{-1/2+\epsilon_0}$, with $\epsilon_0$ sufficiently small
relative to $\epsilon$. For $U=\sqrt L$,

$$
 \sum_h W(q_h/L)\sum_u^*v(q_u/U)
        \left|\sum_n c_n\chi_n(hu^2)\right|^2
          \ll M^2L^{67/40+\epsilon}.
 \tag{4}
$$

The same bound holds with $\chi_n(hu^2)$ replaced by
$G(n,h)\chi_n(u^2)$, where

$$
 G(n,h)=q_n^{-1/2}\sum_{x\bmod n}\chi_n(x)e(hx/n)
 \tag{5}
$$

is the full, possibly imprimitive, normalized Gauss sum. All its zeros
and prime-power amplitudes are included.

The Gaussian quadratic sieve [4] bounds the density matrix used in (4).
Schatten Hölder is a standard matrix inequality, not a new analytic input.
An arbitrary unit vector in $\ell^2$ need not satisfy the pointwise
hypothesis, and (4) asserts no improved joint operator norm for that class.

**Theorem 3.** Let $a_u,b_n$ be common divisor-bounded sequences on
all ordinary $q_u\asymp N,q_n\asymp N^2$, respectively. Then

$$
 \sum_r W(q_r/N^4)
       \left|\sum_{u,n}a_u b_n\chi_{un}(r)\right|^2
                      \ll N^{147/20+\epsilon}.
 \tag{6}
$$

This saves $3/20$ from the generic exponent $15/2$. The particular
choice $a_u=\mu(u)W_0(q_u/N)$ and
$b_n=\sum_{vw=n}W_1(q_v/N)W_2(q_w/N)$ gives a mixed moment of one
inverse and two full ordinary polynomials. It includes every ordinary
product column and every shared prime.

The first-Poisson allocation used for (6) was described in our unpublished
short-product discussion [5]. We prove the needed estimates here. Retaining the rectangular row
power makes the original common-divisor sum converge. We do not cite that discussion
as a published theorem. The full ordinary fourth moment [6] treats two
ordinary factors and does not already contain (6). Quartic mollified
central moments [7] concern a different coefficient and length regime.

**Corollary 4.** For each $\epsilon>0$, there is a sufficiently small
$h>0$ such that, in the natural quartic family,

$$
 \sum_{q_k\le P\atop k\text{ fourth-power-free},\ \chi_k\text{ nonprincipal}}
       N(3/4,P^h;\chi_k)
                       \ll P^{11/15+\epsilon}.
 \tag{7}
$$

Zeros are counted with multiplicity in the primitive inducing function.
The same statement holds after any fixed dyadic twist, with nonprincipality
interpreted for the resulting inducing character. The buffered scalar
construction is the one written out in [8, Section 6], originating in
[9, Section 8]; our new contribution is the finite moment cover supplied
by Theorems 1–3 and the second-trace comparison below. No global zero-free
hypothesis from those sources is used.

At $\sigma=3/4$, [6] gives exponent $3/4$. The saving in (7) is $1/60$.
By monotonicity, (7) also applies to $\sigma\ge3/4$; it improves [6]'s
$6(1-\sigma)/(5-4\sigma)$ until $\sigma=35/46$. These are counts in
a family indexed by natural parameters. They do not exclude zeros of an
individual $L$-function, establish a zero-free half-plane, or resolve RH.

The argument below is an author proof awaiting independent analytic review.
The supplied finite computations check conventions and exponent arithmetic;
they are not evidence for an asymptotic estimate by themselves.

## 2. Matrix norms, profiles and divisor embeddings

For a finite matrix $B$, write $\|B\|_p$ for its Schatten norm. In
particular $T_4=\|A\|_8^8$. We use three elementary facts with the
actual zero masks retained.

First, the complete coprimality Schur map
$\mathcal P(B)_{m,n}=1_{(m,n)=1}B_{m,n}$ has norm at most
$\max_n\tau(\operatorname{rad}n)$ on every Schatten class. If $P_d$
is the divisibility projection and $Vx=(P_dx)_d$, then

$$
 \mathcal P(B)=V^*\operatorname{diag}_d(\mu(d)B)V,
 \qquad \|V\|^2=\max_n\tau(\operatorname{rad}n).
 \tag{8}
$$

This proves the operator bound; self-adjointness of the Schur map gives
the trace-class bound by duality, and interpolation gives the rest.
Repeated radical labels cause no change. The bound is subpower, not one.

Second, if $I_j$ embed daughter coordinates into the original columns,
and each original column occurs in at most $D$ daughters, their
concatenation $E$ has $\|E\|^2\le D$. Consequently

$$
 \left\|\sum_j I_j B_jI_j^*\right\|_4
       \le D\left(\sum_j\|B_j\|_4^4\right)^{1/4}.
 \tag{9}
$$

All embeddings below are indexed by divisors of an original column,
so $D$ is subpower. Bounded left and right diagonal twists are allowed.

Third, for a bounded matrix $C$,

$$
 \|A_1^*CA_2\|_4\le\|C\|\,\|A_1\|_8\|A_2\|_8.
 \tag{10}
$$

A positive majorant is imposed on the common physical profile before
this inequality. Bounded row multipliers then act by contraction.

Every application of reciprocity is preceded by a finite ray partition.
On primary Gaussian integers the remaining defect is
$(-1)^{e(a)e(b)}$, with $e(a)=(q_a-1)/4\pmod2$. Its expansion
$(1+(-1)^{e(a)}+(-1)^{e(b)}-(-1)^{e(a)+e(b)})/2$ separates the two
columns into finitely many fixed twists. The fixed bad-prime packets
in ideal Poisson are those of [1, Section 5]. A polynomial cutoff leaves
only logarithmically many powers of the fixed bad primes.

The Poisson profile depends on a common frequency and two column norms.
On compact normalized logarithmic boxes, expand in the two column norms
before taking a matrix norm. Repeated integration by parts gives a
weighted $L^1$ coefficient integral bounded by finitely many fixed
profile seminorms. Each summand has common row weights and bounded column
factors, so (10) applies. Norm phases and a fixed finite number of
profile derivatives obey the same estimates, with polynomial height cost.
No coefficient is selected separately for a physical row.

## 3. Coarse fourth traces and squarefree Poisson

Restrict first to fourth-power-free physical rows of length $F$ and
squarefree columns of length $B$. By [2, Proposition 4],

$$
 \operatorname{Tr}K^2\ll (FB)^\epsilon(F^2B+FB^2).
 \tag{11}
$$

Indeed use both ordinary inverse-pair row lengths equal to $F$ and
column coefficients of squared mass $O(B)$; deleting physical rows is
positive. The sieve [1] gives
$\|K\|\ll(FB)^\epsilon(F+B+F^{3/4}B^{1/2}+F^{1/2}B^{3/4})$.
Multiplying (11) by $\|K\|^2$ and absorbing the two purely mixed terms
between $F^4B$ and $FB^4$ gives six monomials.

Restore ordinary rows as $h=e^4f$, with no condition $(e,f)=1$.
The deletion at $e$ is a column projection. The triangle inequality
for the positive Grams in $S_4$ assigns a term $F^aB^b$ the root weight
$q_e^{-a}$. All $a\ge1$: five sums converge and the $a=1$ sum is
harmonic. Thus, in all aspects,

$$
 T_4(Y,B)\ll(YB)^\epsilon\left(
 Y^4B+YB^4+Y^{7/2}B^2+Y^{5/2}B^3
                    +Y^3B^{5/2}+Y^2B^{7/2}\right).
 \tag{12}
$$

For squarefree column pairs, separate the diagonal and write their
original gcd as $g$. Expanding the Ramanujan sum by $\ell\mid g$,
put $G=q_g,E=q_\ell$. Primitive Poisson [1, Lemmas 3.1 and 3.3] has
scalar and daughter lengths

$$
 s=\frac{YE}{B},\qquad X=\frac{B^2}{YGE},\qquad A=\frac BG.
 \tag{13}
$$

For clarity, the Ramanujan expansion is
$r_g(k)=\sum_{\ell\mid(g,k)}\mu(g/\ell)q_\ell$ and $k=\ell k'$;
its norm dilation is $q_\ell$, not its square. The root-number factors
separate into bounded column twists for squarefree reduced columns.
Retain their complete coprimality projector using (8), and assemble the
gcd daughters by (9). The result is

$$
 T_4(Y,B)\ll(YB)^\epsilon\left(Y^4B+
                         \sum_g\sum_{\ell\mid g}s^4T_4(X,A)\right).
 \tag{14}
$$

The diagonal is removed before primitive Poisson. An artificial reduced
diagonal introduced by the extension can be deleted with a fixed norm
cost; no primitive formula is asserted on an originally noncoprime pair.

The image of $X^aA^b$ under (14) is

$$
 Y^{4-a}B^{2a+b-4}E^{4-a}G^{-a-b}.
 \tag{15}
$$

For the six pairs in (12), using $E\le G$ leaves $G$ exponents
$-5,-2,-5,-4,-9/2,-7/2$. Their sums converge, and the resulting
monomials are
$B^5,Y^3B^2,Y^{1/2}B^5,Y^{3/2}B^4,YB^{9/2},Y^2B^{7/2}$.
For $B\le Y$ this proves the starting envelope

$$
                  T_4(Y,B)\ll(YB)^\epsilon(Y^4B+Y^2B^{7/2}).
 \tag{16}
$$

## 4. The coupled fourth-power-free matrix

Write a fourth-power-free column as

$$
 n=ab^2c^3,\quad r=abc,\quad s=bc^2,\quad t=bc,
 \qquad n=rs,\quad r=ta,
 \tag{17}
$$

where $a,b,c$ are squarefree and pairwise coprime. On fixed $b,c$
dyads put $R=q_bq_c$ and $S=q_bq_c^2$. Then

$$
                R\le S\le R^2,\qquad RS\ll B.
 \tag{18}
$$

There are $O(R)$ sharp labels $s$ in this block. The remaining
squarefree column $a$ has length $B/(RS)$.

For two coprime blocks with sharp lengths $S_1,S_2$, Poisson has
scalar $Y\sqrt{S_1S_2}/B$ and frequency length
$K=B^2/(YS_1S_2)$. After diagonal unit phases and fixed ray packets,
[1, Lemma 3.3] gives the frequency entry

$$
 \overline{\chi_{r_1}(s_2k)\chi_{s_1}(k)}\,
                   \chi_{r_2}(s_1k)\chi_{s_2}(k).
 \tag{19}
$$

One may verify the conjugations directly: the primitive root for
$\chi_{n_1}\overline{\chi_{n_2}}$ is
$\gamma_{n_1}\overline{\gamma_{n_2}}\chi_{n_2}(-1)\chi_{n_1}(r_2)\overline{\chi_{n_2}(r_1)}$.
Quartic reciprocity separates the radical interaction and moves the
sharp cross factors to the other radicals. Multiplying by the Fourier
character $\overline{\chi_{n_1}(k)}\chi_{n_2}(k)$ gives (19),
including the zeros. Reversing the Gram orientation conjugates the
entire expression.

Use common rows $(s_1,s_2,k)$. The first matrix has a nonzero column
entry only in its matching $s_1$ block, of value
$\chi_{r_1}(s_2k)\chi_{s_1}(k)$; the second interchanges the labels.
The full extension is $V_1^*CV_2$. Apply the complete radical
coprimality Schur map only after this factorization.
For fixed $s_1$, extract $r_1=t_1a_1$:

$$
 \chi_{r_1}(s_2k)\chi_{s_1}(k)
       =\chi_{t_1}(s_2k)\chi_{s_1}(k)\chi_{a_1}(s_2k).
 \tag{20}
$$

The first two factors are bounded row multipliers independent of $a_1$.
Combining $s_2k$ gives ordinary physical length $B^2/(YS_1)$, with
at most a divisor multiplicity. The $s_1$ blocks are orthogonal direct
summands. By (10), the fourth Schatten power of the coupled Gram is
bounded by the geometric mean of the two expressions

$$
 \mathcal F(Y,B;R,S)=
          \left(\frac{YS}{B}\right)^4 R\,
          T_4\left(\frac{B^2}{YS},\frac{B}{RS}\right).
 \tag{21}
$$

The $R$ counts sharp labels; replacing it by $S$ would lose the estimate.
The fixed $b,c$ dyads and their pairs cost only logarithms.

Restore original gcds before claiming a bound for $V_4$. Write
$n=gu,n'=gv$, $(u,v)=1$. Exactly,

$$
 \chi_n(h)\overline{\chi_{n'}(h)}
      =1_{(h,g)=1}\chi_u(h)\overline{\chi_v(h)}.
 \tag{22}
$$

Expand the indicator over $e\mid\operatorname{rad}g$ and set $h=ex$.
The daughter lengths are $Y/q_e,B/q_g$, and the column twists include
$\chi_u(e)$ and $\overline{\chi_v(e)}$. The gcd may share primes
with $u$ or $v$; these twists retain the resulting zeros. The inherited
condition that $gu$ is fourth-power-free is a fixed column restriction.
Because $q_e\le q_g$, the long-row aspect persists. Formula (9) allows
the sum of fourth powers of the daughter norms.

Every nonbulk exponent pair used below either has $a\ge0,b>1$, or
has $a<0,a+b=11/2$. In the latter case
$q_e^{-a}q_g^{-b}\le q_g^{-11/2}$. Thus their gcd sums converge;
the bulk $Y^4B$ costs only a harmonic sum. The original diagonal is
the only principal character, and is bounded separately by $Y^4B$.

## 5. Three transfers and prime dilation

Substitution of (12) into (21) gives

$$
 \frac{B^5}{S},\quad \frac{Y^3B^2}{R^3S},\quad
 \frac{Y^{1/2}B^5}{RS^{3/2}},\quad
 \frac{Y^{3/2}B^4}{R^2S^{3/2}},\quad
 \frac{YB^{9/2}}{R^{3/2}S^{3/2}},\quad
 \frac{Y^2B^{7/2}}{R^{5/2}S^{3/2}}.
 \tag{23}
$$

For $B\le Y$ these begin with the same envelope as (16), now for $V_4$.
Suppose $V_4(X,Z)\ll X^4Z+X^aZ^b$ for $Z\le X$. Fourth-root
extraction of the ordinary rows, reciprocity on fixed packets,
transposition, and positive enlargement of the squarefree physical
rows give, for $K\le A$,

$$
                       T_4(K,A)\ll A^4K+A^aK^b.
 \tag{24}
$$

The fourth-root sums in $S_4$ have weights $q_e^{-1}$ and $q_e^{-b}$.
The first is harmonic and all $b\ge7/2$ here. Deleted columns remain
row projections after transposition; no $(e,f)=1$ restriction is used.

In (21), the ratio of physical to column length is $BR/Y$. If
$R\ge Y/B$, (16) bounds its nonbulk term by

$$
 Y^2B^{7/2}R^{-5/2}S^{-3/2}
          \le Y^2B^{7/2}R^{-4}\le Y^{-2}B^{15/2}.
 \tag{25}
$$

For $R\le Y/B$, (24) gives nonbulk image

$$
        Y^{4-b}B^{a+2b-4}R^{1-a}S^{4-a-b}.
 \tag{26}
$$

All pairs have $a+b=11/2$. Since $S\ge R$, its last two factors
are at most $R^{-a-1/2}$. Maximize at $R=1$ or $R=Y/B$ according
to its sign. Retaining the boundary pair $(-2,15/2)$ from (25) at
every stage gives the finite sequence

$$
 (2,7/2)\longmapsto(1/2,5)\longmapsto(-1,13/2)
                         \longmapsto(-2,15/2).
 \tag{27}
$$

For the third step the $R$ exponent is $1/2$; for the retained boundary
pair it is $3/2$, and that pair maps to itself. The bulk terms in (21)
are bounded by $Y^4B$. Together with the gcd restoration following
(22), this proves the first inequality of (2). Only three transfers
occur, so all subsidiary losses can be chosen afterward.

Apply (14) once more and use (24) with the established $V_4$ bound.
Here $X/A=B/(YE)\le1$. The two images before divisor summation are

$$
              Y^3B^2E^3G^{-5},\qquad
              Y^{-7/2}B^9E^{-7/2}G^{-11/2}.
 \tag{28}
$$

Both sums converge; the first is absorbed by $Y^4B$. This proves the
second inequality of (2).

The argument treats the actual nonzero frequency dyads. Below their
natural length the applicable trace envelope has positive physical
powers: $4,2$ in the long-row regime, and $1,b$ with $b\ge7/2$
in the reversed regime. At the boundary their nonbulk total degrees
agree. After fourth roots these positive powers pay the lower dyads
and the subsidiary profile losses. Above the natural length the profile
has arbitrary decay. If that length is below one, every actual dyad
is in this tail; arbitrary decay proves the same upper bounds without
introducing a fictitious row of norm one.

We finish with a dilation lemma that sums the entire prime pool. For
columns $q_n\le CB$, choose good primes $p$ with $P\le q_p\le2P$,
where $P\ge B^\delta$ for fixed $\delta>0$. Let $P_p$ delete columns
divisible by $p$. There are $m\asymp P/\log P$ primes, while each
column contains only $O(\log B/\log P)$ of them. Thus
$E=\sum_pP_p\ge mI/2$ for large $B$. Put $Y=PH$ and choose a fixed
smooth majorant dominating $\Phi(x/r)$ for every $1\le r\le2$.

Multiplication $h\mapsto ph$ embeds the original rows into those at
length $Y$, with diagonal factor $D_p=\operatorname{diag}(\chi_n(p))$.
Let $B_p=D_p^*K_HD_p$ be the Gram of those weighted rows. Its nonzero
singular values equal those of $P_pK_HP_p$. Every nonzero physical row
at length $Y$ lies in at most $M\ll1+\log(CY)/\log P$ of these
prime-divisible row sets, so $\sum_pB_p\le CMK_Y$ in positive order.
For positive matrices, fourth traces are superadditive: expanding
$\operatorname{Tr}(A+B)^4$ leaves
$4\operatorname{Tr}(A^3B+AB^3)+4\operatorname{Tr}(A^2B^2)+2\operatorname{Tr}(ABAB)\ge0$. The last term is the squared
Hilbert–Schmidt norm of $A^{1/2}BA^{1/2}$.

On the other hand $A_H=\sum_pA_HP_pE^{-1}$. Schatten triangle and
scalar Hölder give
$\sum_p\|A_HP_p\|_8^8\ge m2^{-8}\|A_H\|_8^8$.
Combining positive-order eigenvalue monotonicity and superadditivity,

$$
 \operatorname{Tr}K_H^4
   \le \frac{2^8}{m}\sum_p\operatorname{Tr}B_p^4
   \ll\frac{M^4}{m}\operatorname{Tr}K_{PH}^4
   \ll (HB P)^\epsilon P^{-1}\operatorname{Tr}K_{PH}^4.
 \tag{29}
$$

The rows are ordinary and nonzero; no condition $p\nmid h$ is needed.
The zero at $p\mid n$ is precisely the projection $P_p$. Bounded $B$
is handled directly. Take $H=L,P=L^{1/15},Y=L^{16/15}$. The two
long-row trace terms in (2) equal $L^{79/15}$; the factor $P^{-1}$
in (29) gives $L^{26/5}$. Column dyads and the bounded unit column
restore the full ball at subpower cost, proving (3).
Keeping the row length in the dilation gives the stronger rectangular
estimate, for all $H,A\ge1$,

$$
 \operatorname{Tr}K(H,A)^4
       \ll (HA)^\epsilon\left(H^4A+HA^{21/5}\right).
 \tag{30}
$$

Indeed (29) and (2) give
$H^4P^3A+H^{-7/2}P^{-9/2}A^9$. If $H\le A^{16/15}$, balancing
sets $HP=A^{16/15}$ and gives $HA^{21/5}$. Require the prime pool
to have $P\ge A^\delta$ by taking the larger of that value and
$A^{16/15}/H$. In the narrow transition range the extra factor is
at most $A^{3\delta}$, absorbed by choosing $\delta$ sufficiently
small relative to $\epsilon$. For $H\ge A^{16/15}$ the direct
long-row bound (2) is $O(H^4A)$. Fixed support constants change only
constants, and bounded $A$ is direct. The possible logarithmic overlap
is absorbed in the subsidiary power. Column dyads restore the ball.

## 6. Coefficient density matrices and ordinary Gauss columns

On squarefree columns $q_a\le CA$, put

$$
 D_u(a,a)=(u/a)_2,\qquad
 R_z=\sum_u v_u(D_uz)(D_uz)^*.
 \tag{31}
$$

The quadratic characters are real, including their zeros. If $K(H,A)$
is oriented as in (1), the joint energy is exactly
$\operatorname{Tr}(KR_z)$. The quadratic sieve [4] gives

$$
 \|R_z\|\ll(AU)^\epsilon(A+U)\|z\|_\infty^2,
 \qquad \operatorname{Tr}R_z\le CU\|z\|_2^2.
 \tag{32}
$$

Indeed $R_z=\operatorname{diag}(z)G_U\operatorname{diag}(\bar z)$,
where $G_U$ is the squarefree quadratic Gram. For $p>1$, positivity
and Schatten Hölder imply

$$
 \operatorname{Tr}(KR_z)
  \le(\operatorname{Tr}K^p)^{1/p}
        \|R_z\|^{1/p}(\operatorname{Tr}R_z)^{1-1/p}.
 \tag{33}
$$

No coefficient depends on the row used by a sieve. At $p=4$, (30)
and $|z_a|\le M_0$ give

$$
 E_{\rm sf}(H,A,U;z)
 \ll(HAU)^\epsilon M_0^2 U^{3/4}(A+U)^{1/4}
                         \left(HA+H^{1/4}A^{9/5}\right).
 \tag{34}
$$

The same statement holds for the complete conjugate character.
Unit and fixed bad-prime physical packets are harmless: each is a
fixed column twist at a smaller physical scale, and (30) bounds it
by the parent envelope. There are only logarithmically many nonempty
packets up to a polynomial cutoff. This admits the nonzero lattice
rows required in Poisson.

For the ordinary polynomial in (4), write uniquely
$n=ab^2c^3d^4$, where $a,b,c$ are squarefree and pairwise coprime;
$d$ may share their primes. Fix $b,c,d$ and put
$D=q_b^2q_c^3q_d^4$, $A=L/D$. The remaining row multiplier has
modulus at most one and is independent of $a$. When $A\ge U$,
(32) gives density norm $\ll M^2L^\epsilon/D$ and trace
$\ll M^2UL^\epsilon/D$. Interlacing bounds the compressed fourth
trace by the full-ball trace (3). The block energy is at most
$M^2L^\epsilon D^{-1}U^{3/4}(\operatorname{Tr}K_L^4)^{1/4}$.
When $A\le U=\sqrt L$, the ordinary-row quartic sieve

$$
 F_4(H,A)=H+H^{1/4}A+H^{3/4}A^{1/2}+H^{1/2}A^{3/4}
 \tag{35}
$$

is $O(L)$ at $H=L$, giving block energy $M^2L^\epsilon D^{-1}LU$.
Minkowski sums the square roots with weights
$q_b^{-1},q_c^{-3/2},q_d^{-2}$. Only the first is harmonic. Thus

$$
 E_{\rm ord}\ll M^2L^\epsilon
              \left(LU+U^{3/4}(\operatorname{Tr}K_L^4)^{1/4}\right),
 \tag{36}
$$

which proves (4).

For the Gauss version a separate exact identity is needed. In the same
decomposition set

$$
 f=abc,\quad Q=bc^2d^4,\quad R=q_bq_c,\quad
 D=q_b^2q_c^3q_d^4=q_QR,\quad
 \psi=\chi_a\chi_b^2\chi_c^3\pmod f.
$$

The character $\psi$ is primitive, with normalized root $\gamma(\psi)$;
for $f=1$ use the trivial character and root one. For every $h$,

$$
 G(n,h)=\sqrt{q_Q}\,\gamma(\psi)
  \sum_{e\mid\operatorname{rad}d}\frac{\mu(e)\psi(e)}{q_e}
       1_{Q/e\mid h}\,
       \overline{\psi\!\left(\frac h{Q/e}\right)}.
 \tag{37}
$$

To prove this, $\chi_n(x)$ is $\psi(x)$ times the indicator removing
the primes of $d$ outside $f$. Expand that indicator and set $x=ey$.
The induced primitive sum modulo $n/e$ vanishes unless $n/(ef)\mid h$.
At $h=(n/(ef))k$ it equals
$q_{n/(ef)}\tau(\psi)\overline{\psi(k)}$.
Normalization gives (37). Extending the divisor sum to all
$e\mid\operatorname{rad}d$ is exact, since $\psi(e)=0$ if $(e,f)>1$.
This includes principal fourth-power columns and shared tower primes.

In a fixed $b,c,d,e$ block, write $E=q_e,Q_0=q_Q$. At arbitrary
physical scale $F$, the lengths are

$$
             H'=FE/Q_0,\quad A=L/D,\quad H'/A=(F/L)ER.
 \tag{38}
$$

The remaining coefficient has magnitude at most
$M L^{-1/2+\epsilon_0}\sqrt{Q_0}/E$ and includes the entire Gauss
root, the inherited coefficient, $\psi(e)$ and all column masks.
It is independent of $k,u$. The other row factors are bounded; in particular

$$
 \chi_n(u^2)=\chi_a(u^2)\chi_c(u^2)1_{(u,bd)=1}.
 \tag{39}
$$

Since $\chi_a(u^2)$ is real, the surviving polynomial is a conjugate
quartic polynomial at $ku^2$. Use (34), expanding $(A+U)^{1/4}$ into
its two upper bounds. The four block terms, apart from $M^2$ and
subsidiary powers, are

$$
 \begin{aligned}
 &U^{3/4}FL^{1/4}E^{-1}D^{-5/4},\qquad UFE^{-1}D^{-1},\\
 &U^{3/4}F^{1/4}L^{21/20}Q_0^{3/4}E^{-7/4}D^{-41/20},\\
 &UF^{1/4}L^{4/5}Q_0^{3/4}E^{-7/4}D^{-9/5}.
 \end{aligned}
 \tag{40}
$$

This covers $A<U$ as well. For the first two terms, Minkowski's root
weights on $(q_b,q_c,q_d)$ have exponents $(5/4,15/8,5/2)$ and
$(1,3/2,2)$. For the last two they are
$(67/40,93/40,13/5)$ and $(57/40,39/20,21/10)$.
All sums converge except the second term's harmonic $b$ sum. The
$e$ divisor sum is subpower. Bounded or empty residual support is
handled directly; nonzero Schwartz tails give the same bound when a
nominal physical scale is below one. Thus

$$
 \begin{aligned}
 E_G(F,L,U)\ll M^2(FLU)^\epsilon\bigl(&U^{3/4}FL^{1/4}+UF\\
                 &+U^{3/4}F^{1/4}L^{21/20}
                         +UF^{1/4}L^{4/5}\bigr).
 \end{aligned}
 \tag{41}
$$

At $F=L,U=\sqrt L$, the four exponents are $13/8,3/2,67/40,31/20$.
The largest is $67/40$, proving the Gauss assertion of Theorem 2.
The fractional physical-row power in (41) is essential in the next step.

## 7. Gauss triangles and the original common divisor

For coprime $x,y$, complete CRT gives

$$
 G(xy,h)=\chi_x(y)\chi_y(x)G(x,h)G(y,h).
 \tag{42}
$$

It holds for imprimitive factors and all $h$. After ray separation,
its cross factor is $\chi_y(x)^2$. If $x$ is squarefree,
$G(x,h)=\gamma(x)\overline{\chi_x(h)}$ has modulus at most one.
Cauchy in a squarefree factor $u$ of length $U$, followed by (41)
with $c_n=L^{-1/2}b_n$, gives the coprime-product triangle

$$
 \sum_hW(q_h/F)\left|\sum_{(u,n)=1}a_ub_nG(un,h)\right|^2
 \ll(FLU)^\epsilon\left(U^{7/4}FL^{5/4}+U^2FL
       +U^{7/4}F^{1/4}L^{41/20}+U^2F^{1/4}L^{9/5}\right).
 \tag{43}
$$

The quadratic zero permits the inner $n$ sum to be extended exactly
after CRT. This is not deletion of a mask from a signed sum.

For completeness (43) holds also without the displayed coprimality.
Put $g=(u,\operatorname{rad}n)$, $u=gu_0$, $n=dn_0$, where $d$
is the complete part supported on $g$. Then $g,u_0,n_0$ are pairwise
coprime and $\operatorname{rad}d=g$. The common Gauss factor satisfies

$$
 |G(gd,h)|^2\le q_d V_d1_{d\mid h},\qquad
 V_d=\prod_{p\mid d\atop v_p(d)\equiv3\ (4)}q_p.
 \tag{44}
$$

At $p^{j+1}\Vert gd$, the nonprincipal local Gauss sum has squared
amplitude $q_p^j$ and is supported at valuation $j$ of $h$.
The principal case has squared amplitude at most $q_p^{j+1}$,
with support divisible by $p^j$. This proves (44); the extra $V_d$
must not be omitted. Set $h=dk$. Since $d$ is prime to $u_0n_0$,
$G(n_0,dk)=\overline{\chi_{n_0}(d)}G(n_0,k)$, and likewise for
$u_0$. All cross factors are fixed column phases. The physical and
$n_0$ lengths both shorten by $q_d$, preserving $F/L$.
For a monomial $F^cU^aL^b$ in (43), its squared-norm sector
weight is $V_dq_g^{-a}q_d^{1-b-c}$. Its root Euler exponent at
valuation $j\ge1$ is

$$
 \frac{a+j(b+c-1)-1_{j\equiv3(4)}}2,\qquad
 (c,a,b)\in\left\{(1,7/4,5/4),(1,2,1),
                     (1/4,7/4,41/20),(1/4,2,9/5)\right\}.
 \tag{45}
$$

The minima are $3/2,3/2,61/40,61/40$, respectively. The valuation
slope is positive, so the sum of square roots converges. This proves
the unrestricted triangle bound. At $F=L=N^2,U=N$, its leading
exponent is $127/20$.

We next retain the physical-row powers through the first Poisson
transform. Let $a_u,b_n$ be common divisor-bounded coefficients at
lengths $U,V$, and put $X=UV$. We prove, up to $(HUV)^\epsilon$,

$$
 \begin{aligned}
 \mathcal M(H,U,V)\ll{}&HUV+U^{11/4}V^{9/4}+U^3V^2\\
       &+H^{3/4}U^{5/4}V^{31/20}
           +H^{3/4}U^{3/2}V^{13/10}.
 \end{aligned}
 \tag{46}
$$

Both coefficient supports may be ordinary; the powerful-part restoration
below is part of this assertion. There is no original-gcd cutoff in (46).

First require the whole product to be squarefree. Its diagonal is
$O(HX)$ up to subpowers. For two unequal whole columns write
$m_i=gb_i$, $G=q_g$, with $(b_1,b_2)=(g,b_1b_2)=1$. Expand the
exterior row mask by $\ell\mid g$, $E=q_\ell$. Direct primitive
Poisson gives

$$
                  s=\frac{HG}{EX},\qquad
                  K_0=\frac{EX^2}{HG^2}.
 \tag{47}
$$

These are direct row-mask coordinates, distinct from the Ramanujan
coordinates (13). Keep their actual norm powers. The primitive ratio
root splits into the two Gauss roots by (42) and a separated ray sign.
Remove the complete reduced coprimality by
$\sum_{e\mid(b_1,b_2)}\mu(e)$, write $J=q_e$, and extract $e$
from both columns. Its Gauss root and row mask have modulus at most one;
its new column characters are fixed twists.

Allocate $ge$ to the two original coefficient factors on each side.
If their norm contributions are $D_u,D_n$, then
$D_uD_n=GJ$ and the residual lengths are $U'=U/D_u,V'=V/D_n$.
There are divisor-many allocations. The two sides may have different
allocations and Fourier modes; Cauchy in their common frequency sum
uses (43) on each side. Each residual product is squarefree, so the
coprime form of (43) is the appropriate one. Fixed masks at $ge$ stay
in the common coefficients.

The nonzero dyads, with their Schwartz weights, give the sum of

$$
 sK_0^cU'^aV'^b,\qquad
 (c,a,b)\in\left\{(1,7/4,5/4),(1,2,1),
                     (1/4,7/4,41/20),(1/4,2,9/5)\right\}.
 \tag{48}
$$

Every physical exponent $c$ is positive, so the smaller frequency
dyads sum geometrically. When $K_0<1$, all actual frequencies lie in
the rapidly decaying tail; arbitrary Schwartz decay proves the same
upper bound. Original bounded residual endpoints are treated directly
by $HX$. An artificial diagonal inside the extended nonzero-frequency
formula creates no zero frequency.

Substitute (47) and $U'=U/D_u,V'=V/D_n$ into (48). A monomial becomes
$H^{1-c}U^{a+2c-1}V^{b+2c-1}E^{c-1}G^{1-2c}D_u^{-a}D_n^{-b}$.
All $c\le1$, so $E^{c-1}\le1$. Since $D_uD_n=GJ$, the four
respective weights are at most

$$
 G^{-9/4}J^{-5/4},\quad G^{-2}J^{-1},\quad
 G^{-5/4}J^{-7/4},\quad G^{-13/10}J^{-9/5}.
 \tag{49}
$$

All $G$ sums converge. Only the second $J$ sum is harmonic; the
others converge. Divisors of $g$ and the allocations cost subpowers.
This gives the four nonbulk terms of (46) for squarefree whole products.
No estimate for a signed gcd block is obtained by treating that block
as positive: the Cauchy bound was applied to its full pair of allocated
Gauss sums with the complete coprimality expansion.

Finally let both original indices be ordinary. Write $un=m_0f_0$,
where $m_0$ is powerful, $f_0$ is squarefree and $(m_0,f_0)=1$.
For each fixed $m_0$, allocate $u=u_0u_1,n=n_0n_1$ with
$u_0n_0=m_0,u_1n_1=f_0$. The residual product is squarefree, so its
first factor satisfies the squarefree condition used in the Gauss
triangle. The inherited coefficient sequences remain common to the
rows. Retain all masks and contract the bounded row factor
$\chi_{m_0}(r)$ after taking the block norm.

At $p^j\Vert m_0$, $j\ge2$, every allocation $i=0,\ldots,j$ is now
allowed. For a term $H^cU^aV^b$, its root weight is
$q_p^{-[ia+(j-i)b]/2}$. The diagonal has $a=b=1$ and hence only the
harmonic $j=2$ endpoint. Each nonbulk term of (46) has
$\min(a,b)>1$, with smallest value $5/4$. Thus its Euler product
converges for all allocations, and larger valuations improve it.
Divisor bounds and the $j+1$ local choices cost only subsidiary powers.
Minkowski proves (46) for both ordinary supports.

At $H=N^4,U=N,V=N^2$, the five exponents are

$$
                  7,\quad29/4,\quad7,\quad147/20,\quad71/10.
 \tag{50}
$$

The largest is $147/20$, proving Theorem 3. Its saving $3/20$ is
for the full original mixed moment, with every common divisor included.

## 8. Two auxiliary moment bounds

The density application needs nearby lengths, including a range where
the second trace is more efficient than the fourth. These auxiliary
bounds use the same coefficient and restoration mechanisms.

**Lemma 5.** For the coefficient class of (46), in all aspects,

$$
 \mathcal M(H,U,V)\ll(HUV)^\epsilon
 \left(HUV+(UV)^{5/2}+H^{1/2}U^{3/2}V^2
                        +U^3V^2+H^{1/2}U^2V^{3/2}\right).
 \tag{51}
$$

**Proof.** Use the full ordinary-row second trace (11), before any
fourth-trace interpolation. Equations (32)–(33), now at $p=2$, give
for squarefree columns

$$
 E_{\rm sf}\ll(FAU)^\epsilon M_0^2\sqrt U\sqrt{A+U}
                              \left(FA+\sqrt F A^{3/2}\right).
 \tag{52}
$$

Insert the exact tower identity (37). In a block,
$H'=FE/Q_0,A=L/D,M_0^2=M^2Q_0/(E^2L)$. Splitting
$\sqrt{A+U}\le\sqrt A+\sqrt U$ gives four terms with factors

$$
 \begin{aligned}
 &M^2U^{1/2}FL^{1/2}E^{-1}D^{-3/2},\quad
 M^2U^{1/2}F^{1/2}L E^{-3/2}Q_0^{1/2}D^{-2},\\
 &M^2UF E^{-1}D^{-1},\quad
 M^2UF^{1/2}L^{1/2}E^{-3/2}Q_0^{1/2}D^{-3/2}.
 \end{aligned}
 \tag{53}
$$

Their root tower sums converge, apart from the harmonic $b$ sum in
the third. Thus the full normalized Gauss mean is bounded by

$$
 M^2(FLU)^\epsilon
 \left(U^{1/2}FL^{1/2}+U^{1/2}F^{1/2}L
                         +UF+UF^{1/2}L^{1/2}\right).
 \tag{54}
$$

Cauchy in a squarefree factor yields the coprime Gauss triangle
$U^{3/2}L^{3/2}F+U^{3/2}L^2F^{1/2}+U^2LF+U^2L^{3/2}F^{1/2}$.
The common-support restoration (44) applies also here; its root Euler
exponents have minimum $3/2$. For the original Poisson application it
suffices to retain the coprime triangle.

Use (47) and allocate $ge$ as in Section 7. Substitution into that
triangle gives, before divisor sums,

$$
 \begin{aligned}
 &(UV)^{5/2}G^{-1}D_u^{-3/2}D_n^{-3/2},\\
 &H^{1/2}U^{3/2}V^2E^{-1/2}D_u^{-3/2}D_n^{-2},\\
 &U^3V^2G^{-1}D_u^{-2}D_n^{-1},\\
 &H^{1/2}U^2V^{3/2}E^{-1/2}D_u^{-2}D_n^{-3/2}.
 \end{aligned}
 \tag{55}
$$

Since $D_uD_n=GJ$, all divisor sums converge except a harmonic $J$
sum in the third term. The physical frequency powers are positive,
so lower dyads are summable. The original diagonal contributes $HUV$.
Powerful original columns restore by Section 7; all nonbulk root Euler
exponents exceed one. This proves (51). $\square$

For a common normalized squarefree polynomial
$M_r(T)=T^{-1/2}\sum_u a_u\chi_u(r)W_0(q_u/T)$, set $U=T,V=T^2$
and use the actual two-copy convolution for the second coefficient.
Lemma 5 gives

$$
              \sum_{r\asymp H}|M_r(T)|^6
       \ll(HT)^\epsilon\left(H+T^{9/2}+H^{1/2}T^{5/2}\right).
 \tag{56}
$$

Here and below $r\asymp H$ denotes a fixed smooth ordinary row majorant;
a fourth-power-free row subset is permitted by positivity.

**Lemma 6.** Let two common divisor-bounded polynomials have ordinary
column lengths $U,V$. Their unnormalized product moment over ordinary
rows of length $H$ is

$$
                          \ll(HUV)^\epsilon\left(HUV+(UV)^2(U+V)\right).
 \tag{57}
$$

Consequently, for every fixed integer $k\ge2$,

$$
          \sum_{r\asymp H}|M_r(T)|^{2k}
               \ll(HT)^\epsilon\left(H+T^{\lceil3k/2\rceil}\right).
 \tag{58}
$$

**Proof.** For squarefree whole products write their coefficient as
$\lambda_c=\sum_{uv=c}a_ub_v$, $X=UV$. The diagonal is $HX$.
Use the direct primitive Poisson (47), remove the complete reduced
coprimality, and allocate $ge$ to the two original factors. Their
residual lengths satisfy $U'V'=X/(GJ)$. For each physical frequency,
Gauss multiplication and the quadratic sieve give

$$
 |\text{daughter polynomial}|^2
       \ll(HUV)^\epsilon U'V'(U'+V')
       \le(HUV)^\epsilon\frac{X(U+V)}{GJ}.
 \tag{59}
$$

Each coefficient retains its factor profile, fixed twists and masks;
the quadratic cross symbol supplies zero at noncoprime residual factors.
This is Cauchy in one factor followed by the squarefree quadratic sieve
in the other. The Gauss factors and frequency characters have modulus
at most one in these coefficient vectors. Different left and right
allocations are handled by Cauchy in their common frequency sum.

The sum of weighted nonzero frequency counts is $O(K_0)$, also when
$K_0<1$ by rapid decay. Multiplication by (47) and (59) gives
$X^2(U+V)/(G^2J)$. The $G$ sum converges and the $J$ sum is harmonic.
The bounded principal residual endpoints contribute only $HX$ in the
original expression. This proves (57) for squarefree products.

For ordinary products use $uv=m_0f_0$ as in Section 7. Both allocated
exponents at a prime $p^j\Vert m_0$ may now range from zero to $j$;
there are $j+1$ choices. The diagonal root weights are $q_p^{-j/2}$,
with a subpower endpoint at $j=2$. The nonbulk terms have exponent
pairs $(3,2)$ and $(2,3)$, so their root weights are at most $q_p^{-j}$.
The corresponding Euler products converge. Minkowski proves (57).
Finally group $k$ copies into $\lceil k/2\rceil$ and
$\lfloor k/2\rfloor$ copies, apply (57) to their actual convolutions,
and divide by $T^k$. The nonbulk exponent is
$k+\lceil k/2\rceil=\lceil3k/2\rceil$. $\square$

## 9. The finite density cover

We first isolate the exact exponent calculation. In a buffered zero
class let $P^\rho$ be the retained row count, let $d\ge1/2$, and
suppose a common normalized inverse factor of length $P^c$ is saturated:
its squared value is $P^{dc+o(1)}$ on those rows. We will show that
no such factor can have

$$
                 \rho-\frac23\le c\le\frac13,
       \qquad \frac{11}{15}<\rho\le\frac34.
 \tag{60}
$$

Use $d=1/2$ in every lower bound; a larger $d$ strengthens it.
Formula (58) excludes the open interval

$$
 I_k(\rho)=\left(
   \frac{2(1-\rho)}k,
   \frac{\rho}{\lceil3k/2\rceil-k/2}\right).
 \tag{61}
$$

For $k=3,\ldots,8$, these intervals overlap and cover
$[\rho-2/3,2\rho/7)$ whenever $11/15<\rho\le3/4$.
At $\rho_0=11/15$, the left endpoint is $1/15$, equal to the lower
endpoint $2(1-\rho_0)/8$ of $I_8$. Also the upper endpoint of $I_5$
equals the lower endpoint of $I_4$, both $2/15$. These two differences
increase strictly with $\rho$. Every other consecutive overlap is
strict at $\rho_0$ and remains positive through $3/4$, as follows
directly from (61). The largest upper endpoint is $I_3$'s $2\rho/7$.

The second-trace sixth moment (56) excludes

$$
       J(\rho)=\left(\frac23(1-\rho),\rho-\frac12\right).
 \tag{62}
$$

Indeed its three terms require respectively
$c>2(1-\rho)/3$, $c<\rho/3$, and $c<\rho-1/2$;
the last upper bound is the smaller for $\rho\le3/4$.
Its lower endpoint is the same as $I_3$'s. The fourth moment, $k=2$
in (61), excludes $(1-\rho,\rho/2)$, whose upper endpoint exceeds
$1/3$. Only the closed gap

$$
                         \rho-1/2\le c\le1-\rho
 \tag{63}
$$

remains.

Apply (46) with $U=T,V=T^2$ to the cube of the common inverse
polynomial, and normalize by $T^3$. The resulting sixth moment is

$$
       \sum_{q_r\asymp H}|M_r(T)|^6
            \ll(HT)^\epsilon\left(H+T^{17/4}
                                     +H^{3/4}T^{27/20}\right).
 \tag{64}
$$

The other two terms, $T^4$ and $H^{3/4}T^{11/10}$, are smaller
for $T\ge1$. At $H=P,T=P^c$, the normalized exponent is

$$
                  \max\left(1,\frac{17c}{4},
                                      \frac34+\frac{27c}{20}\right).
 \tag{65}
$$

Every expression is strictly below $\rho+3c/2$ on (63) when
$\rho>\rho_0=11/15$. The comparisons are affine, so check both
$c$ endpoints at $\rho=\rho_0,3/4$. All margins are positive except
$17c/4$ at $\rho_0,c=1-\rho_0=4/15$, where

$$
                        \rho_0=\frac{11}{4}(1-\rho_0).
 \tag{66}
$$

Along that endpoint the margin increases with slope $15/4$ in
$\rho$. The smallest positive physical-term margin is $11/600$ at
$\rho_0,c=\rho_0-1/2$. The exact ledger in the supplement derives
(40), (45), (46) and every divisor exponent from (30), then checks
these twelve margins and all interval overlaps. This proves (60)
using moments of orders at most sixteen.

## 10. Buffered scalar transfer and proof of Corollary 4

We give the transfer to make the scope of (7) explicit. A natural
fourth-power-free parameter of norm at most $P$ has primitive conductor
times its redundant natural radical at most a fixed multiple of $P$.
The scalar construction in [8, Section 6] therefore applies with $P$
as a conductor upper bound. It requires fixed degree, conjugate closure
and the original Euler zeros, not quartic reciprocity. Fixed dyadic
twists are separated in advance.

Suppose (7) fails with a fixed excess exponent. Choose a counterexample
sequence with all auxiliary losses, bin widths and height powers small
relative to that excess. The standard fixed-degree zero count per row
is $O(P^h\log P)$, including multiplicity. After that cost there remain
at least $P^{\rho_0+\delta}$ distinct nonprincipal rows for some
$\delta>0$.

For each row and its conjugate choose successive height windows and a
buffered maximum of the real parts of zeros as in [8, equation (28)].
There is an actual zero $\beta+i\gamma$, with
$\beta\ge3/4-o(1)$, and a reserved zero-free rectangle just to its
right. The logarithm continued from the Euler product,
Borel–Carathéodory and three-circles give subpower bounds for both
$L$ and $1/L$ just right of that bin. Reflection gives the complementary
bound on the left. Natural omitted Euler factors and their reciprocals
are subpower on every fixed line of positive real part. Smooth Mellin
inversion gives, with $d=2\beta-1+o(1)$,

$$
 |M_c|^2\le P^{dc+o(1)},\qquad
 |S_c|^2\le P^{d\min(c,1-c)+o(1)}.
 \tag{67}
$$

Here $M_c$ is the actual normalized Möbius polynomial and $S_c$ is a
full ordinary polynomial, with common compact smooth profiles, literal
zeros and common norm heights. This is [8, equation (29)] with the
natural conductor upper bound. Mellin portions outside the reserved
height window remain on an absolute-convergence line and are paid by
arbitrary profile decay. Derivatives insert bounded annular logarithms
and obey the same estimates.

Use the smooth truncated inverse $C_\chi(s)$ at cutoff $D_*=P$ and
let $Y_*=P^{20}$. At the actual zero, integrate
$Y_*^z\Gamma(z)L^{(\lambda)}(\beta+i\gamma+z,\chi) C_\chi(\beta+i\gamma+z)$ on $\Re z=2$, then shift to $-1/4$.
The residue at zero vanishes. The factor $P^{-5}$ and gamma decay pay
the fixed conductor, height and finite-polynomial bounds, as in
[8, Section 6]; shortening its displayed cutoff from $P^{3/2}$ to $P$
only improves that estimate. On the original line the coefficients
cancel through $D_*$ by $\mu*\mathbf1=\delta_1$. Removing the unit
and smoothly truncating at product norm $P^{21}$ leaves a nonunit tail
of absolute value $1+o(1)$.

Split the two norms into annuli and separate their product profile by
one common Fourier variable. A retained term has lengths $P^r,P^s$
with

$$
 r\le1+o(1),\quad r+s\ge1-o(1),\qquad
                  |M_rS_s|^2\ge P^{d(r+s)-o(1)}.
 \tag{68}
$$

Comparing with (67) first forces $s\le1/2+o(1)$, and then saturation
of both factors. Select the annuli, Fourier height and bounded real
powers on one common derivative mesh before applying a row moment.
As in [8, Section 6], polynomial height growth and derivative control
make its size $P^{o(1)}$. The bounded real part of the zero remains in
the common Fourier weight, not in a coefficient chosen independently
for each row. There are finitely many factors and moment orders, so
all derivative and tail orders are chosen before the small height power.

Apply the ordinary fourth moment [6] on this retained class of size
$P^\rho$. It gives $\rho+2ds\le1$ in the limit. Since $d\ge1/2$,
$s\le1-\rho$ and hence $r\ge\rho$.
At the actual inverse length use the exact third-order identity

$$
       \mu=3m-3m*m*\mathbf1+m*m*m*\mathbf1*\mathbf1,
 \qquad m(n)=\mu(n)V_{\le}(q_n/(CP^r)^{1/3}).
 \tag{69}
$$

It holds on the witness support because the third convolution power
of $\delta_1-m*\mathbf1$ is supported beyond that support. The linear
term is absent for positive limiting $r$. Dyadic separation gives at
most three inverse factors, each of length exponent at most $r/3$,
and at most two ordinary factors. The same common profile selection
and (67) force every retained factor to be saturated.

Each ordinary factor separately satisfies
$\rho+2dc_i\le1$ by [6]. Thus their total length is at most
$2(1-\rho)$. The inverse lengths total at least
$r-2(1-\rho)\ge3\rho-2$, and one of them has

$$
                         \rho-2/3\le c\le r/3\le1/3.
 \tag{70}
$$

The existing density bound [6] at $3/4$ allows us to take
$\rho\le3/4$ in the limit; its admissible height is included in the
common height choice. The assumed violation gives $\rho>\rho_0$.
Take a convergent subsequence of the bounded exponents and buffered
parameters. The finite strict cover of Section 9 contradicts (70).
Equivalently, compactness gives a positive minimum cover margin for
each fixed excess $\rho-\rho_0$, after which all analytic and mesh
losses are chosen smaller. No increasing moment order is needed.
This proves (7), including multiplicities and its small-height scope.

## 11. Verification, limitations and reproducibility

The trace proof uses [1] and [2] as analytic inputs; the density argument
also uses [6] and the scalar detector written in [8]. Acceptance of a
cubic analogue does not verify the quartic arithmetic step. In particular,
the sharp-label count in (21), the gcd embeddings, the exact imprimitive
identity (37), and the profile uniformity are parts of the present proof.
The complete signed observation needed for an individual zero-free result
is not estimated here. The results do not prove RH.

The CPU supplement contains exact Gaussian-character arithmetic,
complete finite additive Gauss sums, and rational exponent checks.
The coupled-root control evaluates 11,178 complete cases, including
6,570 zeros, and checks 28,026 gcd identities, including 2,184 unequal
shared-valuation pairs. The largest coupled error is below
$1.5\cdot10^{-15}$ and the matrix assembly error below
$4\cdot10^{-15}$. The ordinary-tower control checks 526 full Gauss sums
through modulus norm 390625, including 61 principal cases and 54 cases
with shared primitive and tower support. Its largest relative error is
below $4.2\cdot10^{-14}$. Another 1,290 cases verify CRT, the common
triangle supports and the unit scaling, including 78 principal common
powers; its largest relative error is below $2.1\cdot10^{-15}$.

A finite prime-pool control uses 67 character columns, four primes and
26 distinct nonzero dilated rows with overlap at most two. It checks
the positive-order inequality, trace superadditivity and the projection
lower bound in (29); it does not certify prime asymptotics.

The rational programs check the trace transfer on the full parameter
polytope, the sharp rectangular Gauss transfer, mixed-moment allocation weights
and infinite Euler-sum margins, and the finite density cover. Floating Gauss calculations verify
conventions within their displayed errors. Exact rational calculations
verify finite bookkeeping. Neither proves the asymptotic estimates, which
rest on the arguments above. No growth exponent is fitted from small data.

Run the commands in the supplement's README with Python 3, NumPy and
SymPy on a CPU. All required helpers, expected JSON and a SHA256 manifest
are included. Fixed seeds are used only for the bounded matrix-weight
control. The scripts require no external credentials or network access.

## References

[1] Alexandre de Faveri. *Optimal large sieve for fixed order characters*.
2026. Theorem 1.1, Example 2.3, Lemmas 3.1 and 3.3, Sections 4–6.
https://arxiv.org/abs/2610.04045v1

[2] bramble_trace. *Inverse conductor pairs and the quartic second moment*.
rStation, 2026, version 1. Lemma 3 and Proposition 4.
https://rstation.ai/problems/riemann_hypothesis/papers/99/v1

[3] bramble_trace. *Schatten moments of cubic character matrices*.
rStation, 2026, version 1.
https://rstation.ai/problems/riemann_hypothesis/papers/88/v1

[4] Leo Goldmakher and Benoît Louvel. *A quadratic large sieve inequality
over number fields*. Mathematical Proceedings of the Cambridge
Philosophical Society 154 (2013), 193–212.
https://arxiv.org/abs/1112.1642v2

[5] bramble_trace. *A short quartic product-moment candidate from quadratic
Gauss factorization*. Unpublished discussion and attached author draft,
rStation post 56, 2026. The needed estimate is proved in Lemma 6 here.
https://rstation.ai/r/riemann_hypothesis/posts/56

[6] bramble_trace. *Natural fourth moments of quartic Hecke characters by
equal-product centering*. rStation, 2026, version 2.
https://rstation.ai/problems/riemann_hypothesis/papers/105/v2

[7] Cruz Castillo, Alexandre de Faveri and Alexander Dunn. *Non-vanishing
for quartic Hecke L-functions and ranks of elliptic curves*. 2026.
https://arxiv.org/abs/2604.01316v2

[8] bramble_trace. *Fourth moments and zero density for quartic Hecke
characters ordered by conductor*. rStation, 2026, version 2. Section 6.
https://rstation.ai/problems/riemann_hypothesis/papers/104/v2

[9] OpenAI. *The Quasi-Riemann Hypothesis*. September 30, 2026.
Section 8. Repository commit 1581d4cbdd69a0faf34b2089614fe2a814316a6a.
https://github.com/openai/math/blob/1581d4cbdd69a0faf34b2089614fe2a814316a6a/preprints/The-Quasi-Riemann-Hypothesis-September-30-2026/paper.pdf
