\documentclass[11pt,a4paper]{article}
\usepackage[T1]{fontenc}
\usepackage[utf8]{inputenc}
\usepackage{lmodern}
\usepackage[margin=28mm]{geometry}
\usepackage{amsmath,amssymb,amsthm,mathtools,bm}
\usepackage{booktabs,array,microtype}
\usepackage{xurl}
\usepackage{amsthm}
\newtheorem{theorem}{Theorem}
\usepackage[hidelinks,pdfusetitle]{hyperref}
\numberwithin{equation}{section}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{maintheorem}{Theorem}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\theoremstyle{remark}
\newtheorem{remark}[theorem]{Remark}
\newcommand{\R}{\mathbb R}
\newcommand{\C}{\mathbb C}
\newcommand{\N}{\mathbb N}
\newcommand{\Z}{\mathbb Z}
\newcommand{\E}{\mathbb E}
\newcommand{\dd}{\,\mathrm d}
\DeclareMathOperator{\supp}{supp}
\DeclareMathOperator{\Res}{Res}
\DeclareMathOperator{\diag}{diag}
\DeclareMathOperator{\vol}{vol}
\newcommand{\repo}{\url{https://github.com/Niklas-Fraccaro/sixth-moment-lean}}
\allowdisplaybreaks[1]
\setlength{\emergencystretch}{2em}
\title{A lower bound for the sixth moment\\of the Riemann zeta function}
\author{Niklas Fraccaro}
\date{}
\begin{document}
\maketitle
\begin{abstract}
We establish a new lower bound for the sixth moment of the
Riemann zeta function on the critical line, improving the
previous constant of 34.4 obtained by Durkan and Page to
41.008, close to the conjectured value of 42 predicted by
the random matrix model of Keating and Snaith.
Our approach extends the two-piece amplifier method by
allowing independent profiles, centered derivatives, and
finite linear combinations of two families of projected
test functions.
\end{abstract}
\smallskip
\noindent\textit{2020 Mathematics Subject Classification.} Primary 11M06.\par
\noindent\textit{Keywords.} Riemann zeta function, sixth moment, amplified moments,
twisted fourth moment.
\section*{Introduction}
Understanding the behavior of the Riemann zeta function
$\zeta(s)$ on the critical line $\Re(s)=\frac12$ is a
central problem in analytic number theory. In particular,
determining the size of its largest values remains a
challenging open problem. A natural approach is to study
the moments of the zeta function, defined by
\[
M_k(T)
:=
\int_0^T
\left|\zeta\left(\frac12+it\right)\right|^{2k}
\,\mathrm{d}t,
\qquad k>0.
\]
Although moments are themselves difficult to analyze,
they provide a more accessible way of studying the
distribution of values of $\zeta$ on the critical line.
Moreover, as $k$ increases, the contribution from large
values of $|\zeta(\frac12+it)|$ becomes increasingly
significant. Thus, higher moments provide valuable
information about the extreme behavior of the zeta
function.
Asymptotic formulas are known for the second and fourth
moments. In particular,
\[
M_1(T)\sim T\log T
\]
and
\[
M_2(T)\sim \frac{1}{2\pi^2}T(\log T)^4
\]
as $T\to\infty$. However, for higher even moments,
no corresponding asymptotic formula has yet been
established. Keating and Snaith~\cite{KS} conjectured that,
for each fixed positive integer $k$,
\[
M_k(T)\sim C_k T(\log T)^{k^2},
\]
where $C_k>0$ is an explicit constant.
The next natural case beyond the fourth moment is
therefore the sixth moment, for which the conjecture
predicts
\[
M_3(T)\sim C_3 T(\log T)^9.
\]
To be more precise, let
\[
a_3 :=
\prod_{p}
\left(1-\frac{1}{p}\right)^4
\left(1+\frac{4}{p}+\frac{1}{p^2}\right),
\]
where the product runs over all primes.
Consistent with the earlier conjecture of Conrey and
Ghosh~\cite{CG}, the Keating--Snaith conjecture~\cite{KS} predicts that
\[
C_3 = \frac{42a_3}{9!}.
\]
\begin{theorem}
\label{thm:main}
We establish the lower bound
\[
\liminf_{T\to\infty}
\frac{M_3(T)}{\frac{a_3}{9!}T(\log T)^9}
> 41.008.
\]
\end{theorem}
\section*{Outline of the Proof}
The main idea is to construct a test function that approximates
$\zeta(s)^3$ well enough to obtain a strong lower bound for its
sixth moment using Cauchy--Schwarz.
We begin with two families of test functions, built from Dirichlet
polynomials and centered derivatives of the zeta function.
Together, these generalize the two-piece amplifier used by
Durkan and Page. The functional equation allows us to project
these functions onto the same real line in $\mathbb{C}$ as
$\zeta(s)^3$. This projection preserves their correlation with
$\zeta(s)^3$ while not increasing their norm. We then take finite
real linear combinations of these projected functions. Applying
Cauchy--Schwarz reduces the problem to finding a vector of mixed
moments and a Gram matrix describing the norms and interactions
of the test functions.
The next step is to evaluate these moments. For the numerator,
the two families require different analytic results. For the
first family, we extend the twisted fourth-moment argument of Bettin, Bui, Li and Radziwiłł~\cite{BBLR} to Dirichlet polynomials of different lengths. This gives the range $3a+5b<2$. For the second family, we use the fourth-moment formula of
Durkan and Page~\cite{DP}, with polynomial lengths below $T^{1/4}$. Although the two calculations are different, we show that their main terms can be expressed using the same kernel, with a reflection in one variable for the second family.
For the denominator, we expand the products of the projected
test functions and use the functional equation to reduce them
to balanced and unbalanced moments involving two zeta factors.
The condition on the polynomial lengths allows us to control
the off-diagonal terms. This gives explicit formulas for all
entries of the Gram matrix, including the mixed terms between
the two families.
Once these formulas are established, the remaining problem is
finite-dimensional. We first work with point evaluations of the
kernels, which are easier to compute exactly, and then approximate
them by admissible smooth or polynomial profiles. This ensures
that the resulting bound applies to genuine test functions.
Finally, we choose 250 explicit tests and evaluate the
corresponding quadratic expressions using exact rational
arithmetic. The resulting quotient is strictly greater than
$41.008$. Applying Cauchy--Schwarz and combining the
estimates over finitely many dyadic intervals removes the
smooth weight and gives the claimed lower bound for the
sixth moment.
\section{Notation and Setup}
\label{sec:setup}
Let $T>1$, $L=\log T$ and $s=\tfrac12+it$. We use the functional
equation in the following convention
\[
\zeta(s)=\chi(s)\zeta(1-s),\qquad
\chi(s)=\pi^{s-1/2}\frac{\Gamma((1-s)/2)}{\Gamma(s/2)}.
\]
On the critical line, $1-s=\bar s$ and $|\chi(s)|=1$, so that
$\zeta(s)^3=\chi(s)^3\overline{\zeta(s)^3}$.
For a real-valued, compactly supported function $f$ with
$\operatorname{supp} f \subseteq [0,a]$, define
\[
A_{f,T}(w):=\sum_{n\geq1} f\!\left(\frac{\log n}{L}\right)n^{-w},
\]
which is a Dirichlet polynomial of length at most $T^a$, since
$f(\log n/L)$ vanishes for $n>T^a$.
For $p\in\R[x]$, we put
\begin{align*}
C_{p,T}(w)
&=p\!\left(\frac12+\frac1L\frac{\mathrm d}{\mathrm dw}\right)\zeta(w)\\
&=\left.p(\partial_z)\bigl[e^{z/2}\zeta(w+z/L)\bigr]\right|_{z=0},
&p^-(x)&=p(-x).
\end{align*}
To simplify notation, we suppress $T$ when fixed. Since the profiles and
polynomial coefficients are real,
\[
A_f(1-s)=\overline{A_f(s)},\qquad C_p(1-s)=\overline{C_p(s)}.
\]
All profiles and test polynomials remain fixed as $T\to\infty$.
We also define
\[
R_T(t,z):=e^z\frac{\chi(s+z/L)}{\chi(s)}.
\]
The functional equation then gives the exact identity
\[
e^{z/2}\zeta(s+z/L)
=\chi(s)R_T(t,z)e^{-z/2}\zeta(1-s-z/L).
\]
\subsection{Test family construction}
To give an idea where our approach starts, we begin with the two-piece amplifier test used by Durkan and
Page~\cite[Section~1.3 and equation~(1.9)]{DP},
\[
\zeta(s)\bigl(A(s)+\chi(s)A(1-s)\bigr)^2.
\]
Expanding the square gives
\begin{align*}
&\zeta(s)\bigl(A(s)+\chi(s)A(1-s)\bigr)^2\\
&\quad=
\zeta(s)A(s)^2
+2\chi(s)\zeta(s)A(s)A(1-s)
+\chi(s)^2\zeta(s)A(1-s)^2.
\end{align*}
Although this expansion contains three terms, there are only
two essentially different types of products: same-side products,
such as $A(s)^2$ and $A(1-s)^2$, and mixed products of the
form $A(s)A(1-s)$.
The two same-side terms are related by the functional equation
of the Riemann zeta function. This suggests working with a
projection that automatically produces the reflected terms.
For a complex-valued function $H(t)$, define
\[
\Pi H(t)
=\frac{H(t)+\chi(s)^3\overline{H(t)}}{2}.
\]
Since $|\chi(s)|=1$ on the critical line, this is pointwise
the orthogonal projection, with respect to the real inner
product $\Re(z\bar w)$, onto the line
\[
\{w\in\C:w=\chi(s)^3\bar w\}.
\]
By the functional equation, this line also contains
$\zeta(s)^3$.
As we said before, $A$ has real coefficients, so that
$\overline{A(s)}=A(1-s)$ on the critical line.
Using
\[
\overline{\zeta(s)}=\chi(s)^{-1}\zeta(s),
\qquad
\overline{\chi(s)}=\chi(s)^{-1},
\]
we obtain
\[
\chi(s)^3\overline{\zeta(s)A(s)^2}
=\chi(s)^2\zeta(s)A(1-s)^2.
\]
Thus the projection of the first same-side term produces
both same-side contributions:
\[
2\Pi\bigl(\zeta(s)A(s)^2\bigr)
=\zeta(s)A(s)^2
+\chi(s)^2\zeta(s)A(1-s)^2.
\]
For the mixed term, we instead have
\[
\chi(s)^3
\overline{\chi(s)\zeta(s)A(s)A(1-s)}
=\chi(s)\zeta(s)A(s)A(1-s).
\]
Hence the mixed term is already invariant under the
reflection, and
\[
2\Pi\bigl(\chi(s)\zeta(s)A(s)A(1-s)\bigr)
=2\chi(s)\zeta(s)A(s)A(1-s).
\]
Consequently, the entire two-piece test can be recovered
by projecting just two terms:
\begin{align*}
&2\Pi\Bigl(
\zeta(s)A(s)^2
+\chi(s)\zeta(s)A(s)A(1-s)
\Bigr)\\
&\qquad=
\zeta(s)\bigl(A(s)+\chi(s)A(1-s)\bigr)^2.
\end{align*}
This observation motivates our two test families.
To further expand the test space, we generalize the zeta factor from $\zeta(s)$ to $C_p(s)$
and allow the two amplifier profiles to vary independently.
We define
\[
U_{p;f,g}(t)
=\chi(s)C_p(s)A_f(s)A_g(1-s),
\]
and
\[
V_{p;f,g}(t)
=C_p(s)A_f(s)A_g(s).
\]
The family $U$ corresponds to the mixed products, whereas
$V$ corresponds to the same-side products.
For the moment calculations we use the following identities:
\begin{equation}\label{eq:projection}
\begin{aligned}
\zeta(s)^3\overline{\Pi H}
&=\Re\bigl(\zeta(s)^3\overline H\bigr),\\
(\Pi H)\overline{\Pi J}
&=\frac12\Re\bigl(H\overline J+\overline{\chi(s)^3}HJ\bigr),\\
|\Pi H|&\leq |H|.
\end{aligned}
\end{equation}
Although both test families are projected using the same operator \(\Pi\), the moment estimates used to analyze their numerators impose different restrictions on the admissible profiles. We therefore specify these restrictions before turning to the analytic estimates.
\subsection{Admissible profiles}
\label{sec:admissible}
For the two test families, we impose the following
conditions on the profiles:
\begin{align*}
U:\quad &f\in C_c^\infty((0,a)),\quad
g\in C_c^\infty((0,b)),
\quad 00$. Whenever the denominator is nonzero,
Cauchy--Schwarz gives
\[
\int_\mathbb R\Phi(t/T)|\zeta(s)|^6\,\dd t
\geq
\frac{
\left|\int_\mathbb R\Phi(t/T)\zeta(s)^3
\overline{F_c(t)}\,\dd t\right|^2
}{
\int_\mathbb R\Phi(t/T)|F_c(t)|^2\,\dd t
}.
\]
The correlation in the numerator is linear in $c$, whereas
the squared norm in the denominator is quadratic in $c$.
Their normalized leading terms will be described by a real
vector $k$ and a real symmetric Gram matrix $G$, giving
rise to the expressions
\[
c^\top k
\qquad\text{and}\qquad
c^\top Gc.
\]
We will determine these coefficients from the individual
correlations and pairwise products of the projected tests
in the following sections.
\section{Analytic Foundations of the Lower Bound}
\label{sec:analytic}
The goal of this section is to determine the numerator vector \(k\) and the denominator Gram matrix \(G\) through explicit limiting kernels, thereby reducing the problem to the optimization of
\[
\frac{(c^\top k)^2}{c^\top Gc}.
\]
We begin with a general nine-edge diagonal limit, whose underlying arithmetic structure also enters the numerator calculation, while the full limiting formula is needed for the denominator. We then derive the numerator asymptotics before turning to the denominator and its associated Gram matrix.
\subsection{Nine-edge Diagonal Limit}
\begin{lemma}[Nine-edge diagonal limit]\label{lem:gram-diagonal}\label{lem:gram-nine-edge}
Let $F$ be bounded, measurable and compactly supported on
$\R_{\geq0}^3\times\R_{\geq0}^3$, and suppose that its discontinuity set has
zero measure for the limiting measure below. Writing
$\log\mathbf n/L=(\log n_1/L,\ldots,\log n_3/L)$, we have
\begin{equation}\label{eq:gram-diagonal}
\frac1{L^9}
\sum_{n_1n_2n_3=m_1m_2m_3}
\frac{F(\log\mathbf n/L,\log\mathbf m/L)}{n_1n_2n_3}
\longrightarrow
a_3\int_{x_{ij}\geq0}
F\left(\left(\sum_jx_{ij}\right)_i,
\left(\sum_ix_{ij}\right)_j\right)
\prod_{i,j=1}^3\dd x_{ij}.
\end{equation}
\end{lemma}
\begin{proof}
Rather than working directly with a general test function $F$,
we first compute the Laplace transforms of the associated
measures. This amounts to considering exponential test functions
of the form
\[
F_{\alpha,\beta}(\mathbf x,\mathbf y)
=\exp\left(-\sum_{i=1}^3\alpha_i x_i
-\sum_{j=1}^3\beta_j y_j\right),
\]
where $\alpha_i,\beta_j>0$.
We first group the diagonal sum according to the common product \(k=n_1n_2n_3=m_1m_2m_3\). For this purpose, define
\[
A_{\mathbf u}(k)=\sum_{n_1n_2n_3=k}\prod_{i=1}^3n_i^{-u_i}.
\]
Since these coefficients are multiplicative, the diagonal Dirichlet series admits an Euler product:
\[
\sum_{\prod_i n_i=\prod_jm_j}\frac{\prod_i n_i^{-u_i}\prod_jm_j^{-v_j}}{\prod_i n_i}=\sum_{k\geq1}\frac{A_{\mathbf u}(k)A_{\mathbf v}(k)}{k}=\prod_p H_p(\mathbf u,\mathbf v),
\]
where the local factor at a prime \(p\) is
\[
H_p(\mathbf u,\mathbf v)=\sum_{r\geq0}p^{-r}\left(\sum_{\substack{a_1+a_2+a_3=r\\a_i\geq0}}p^{-\sum_i u_i a_i}\right)\left(\sum_{\substack{b_1+b_2+b_3=r\\b_j\geq0}}p^{-\sum_j v_j b_j}\right).
\]
The contribution from \(r=0\) is \(1\). For \(r=1\), exactly one exponent on each side equals \(1\), giving \(3\cdot3=9\) possible pairs. Consequently,
\[
H_p(\mathbf u,\mathbf v)=1+\sum_{i,j=1}^3p^{-(1+u_i+v_j)}+O(p^{-2+\varepsilon}),
\]
uniformly for sufficiently small shifts.
These nine first-order terms coincide with those arising from the local Euler factors of the nine zeta functions
\[
\prod_{i,j=1}^3\zeta(1+u_i+v_j).
\]
We therefore define the local remainder
\[
E_p(\mathbf u,\mathbf v)=H_p(\mathbf u,\mathbf v)\prod_{i,j=1}^3\left(1-p^{-(1+u_i+v_j)}\right).
\]
The first-order terms cancel, so
\[
E_p(\mathbf u,\mathbf v)=1+O(p^{-2+\varepsilon}).
\]
Choosing \(\varepsilon<1\), the Euler product defining \(E\) converges locally uniformly near the origin and hence defines a holomorphic function there. Thus
\[
D(\mathbf u,\mathbf v)=E(\mathbf u,\mathbf v)\prod_{i,j=1}^3\zeta(1+u_i+v_j).
\]
Finally, at \(\mathbf u=\mathbf v=\mathbf0\), the number of nonnegative solutions of \(a_1+a_2+a_3=r\) is \(\binom{r+2}{2}\). Hence
\[
\begin{aligned}E_p(\mathbf0,\mathbf0)&=(1-p^{-1})^9\sum_{r\geq0}\binom{r+2}{2}^{\!2}p^{-r}\\&=(1-p^{-1})^4(1+4p^{-1}+p^{-2}).\end{aligned}
\]
Taking the product over all primes gives \(E(\mathbf0,\mathbf0)=a_3\).
Substituting \(u_i=\alpha_i/L\) and \(v_j=\beta_j/L\), with positive real parts, and normalizing by \(L^{-9}\), yields the limiting Laplace transform
\[
\lim_{L\to\infty}L^{-9}D(\alpha/L,\beta/L)
=a_3\prod_{i,j}(\alpha_i+\beta_j)^{-1}.
\]
The limiting Laplace transform can be identified by considering
the map that sends the nine nonnegative variables $x_{ij}$
to their row and column sums. Indeed,
\[
a_3\int_{x_{ij}\geq0}
e^{-\sum_{i,j}(\alpha_i+\beta_j)x_{ij}}
\prod_{i,j}\mathrm{d}x_{ij}
=
a_3\prod_{i,j}(\alpha_i+\beta_j)^{-1}.
\]
Thus, the limit is precisely the Laplace transform of the
pushforward of $a_3$ times Lebesgue measure under this map.
Although the measures involved are not finite, multiplying
them by a fixed positive exponential weight makes them finite.
Their Laplace transforms still converge, so the continuity
theorem for Laplace transforms yields weak convergence of
the weighted measures.
Finally, since $F$ has compact support, the exponential
weight is bounded away from zero on its support.
We can therefore remove the weight and use weak convergence
for bounded measurable functions whose discontinuity sets
have limiting measure zero. This proves the assertion.
\end{proof}
\subsection{Numerator}
\label{sec:numerator}
The numerator is determined by the correlations of the two test families, $U$ and $V$, with $\zeta(s)^3$. Although these correlations lead to different types of twisted moments, their limiting expressions can be described by a single kernel. The purpose of this section is to identify this common kernel and derive the corresponding numerator formulas for both families.
We begin by introducing the kernel and stating the main result. For the $U$-family, we use an asymmetric twisted fourth moment, which allows the two amplifier profiles to be different and to appear on opposite sides of the moment. Its six main terms can then be combined and reduced to a simplex integral. The $V$-family requires a different approach, based on the fourth-moment formula of Durkan–Page. Finally, we establish a reflection identity connecting the two expressions. This shows that both families are governed by the same kernel, with a reflected argument in the $V$-case, and completes the evaluation of the numerator.
\subsubsection{The common numerator kernel and the main result}
\label{sec:num-kernel}
We first introduce this common kernel and state
the corresponding limits, before turning to their derivation
in the following subsections.
Let $X=(X_0,X_1,X_2,X_3)$ be uniformly distributed on the
three-dimensional simplex
\[
X_j\geq 0,\qquad \sum_{j=0}^3 X_j=1.
\]
We normalize the associated expectation by
\[
\E F(X)
=6\int_{\substack{x_1,x_2,x_3\geq0\\
x_1+x_2+x_3\leq1}}
F(1-x_1-x_2-x_3,x_1,x_2,x_3)
\,\dd x_1\dd x_2\dd x_3,
\qquad \E1=1.
\]
For $a,b\geq0$, define
\begin{align}
P(u,v)
&=\frac{uv(1-u)(1-v)(1-u-v)}{12},
\notag\\
q(a,b,t)
&=\left(
-\frac12,\,
a-t-\frac12,\,
\frac12-b+t,\,
\frac12
\right),
\notag\\
K_z(a,b)
&=\int_0^{\min(a,b)}
P(b-t,a-t)\,
\E e^{zq(a,b,t)\cdot X}\,\dd t.
\label{eq:num-kernel}
\end{align}
Here $a$ and $b$ represent logarithmic positions in the
amplifier profiles, rather than fixed support bounds.
For each fixed $a,b$, the function $K_z(a,b)$ is entire
in $z$.
The parameter $z$ allows us to incorporate the polynomial
weights appearing in the numerator. For a real polynomial
$p$, we define
\begin{equation}\label{eq:num-poly-kernel}
\begin{aligned}
K_p(a,b)
&=p(\partial_z)K_z(a,b)\big|_{z=0}\\
&=\int_0^{\min(a,b)}
P(b-t,a-t)\,
\E p\bigl(q(a,b,t)\cdot X\bigr)\,\dd t.
\end{aligned}
\end{equation}
Thus, the kernels associated with different polynomial
weights are obtained from the same generating function
$K_z$.
We can now state the numerator limits for both test
families in a common form.
\begin{proposition}[Numerator limits]\label{prop:numerator}
For an admissible $U$- or $V$-test and a real polynomial $p$
of degree at most four, the corresponding limiting numerator
coefficients are given by
\begin{align}
k_U(p;f,g)
&=\int_0^\infty\int_0^\infty
f(a)g(b)K_p(a,b)\,\dd a\dd b,
\label{eq:num-u-limit}\\
k_V(p;f,g)
&=\int_0^\eta\int_0^\eta
f(a)g(b)K_p(a,1-b)\,\dd a\dd b.
\label{eq:num-v-limit}
\end{align}
No additional projection factor appears in these formulas.
\end{proposition}
\subsubsection{An asymmetric twisted fourth moment}
\label{sec:twisted}
The $U$-numerator contains one Dirichlet polynomial on each side of the
critical line. Their lengths need not be equal, and imposing the same
upper bound on both would unnecessarily restrict the available tests.
We therefore retain the two lengths separately in the argument of
Bettin--Bui--Li--Radziwi\l\l\ \cite{BBLR}.
The main terms remain unchanged; the point is to preserve the finer
length dependence in their off-diagonal estimates.
We first specify the moment and its complete main term. We then collect
the ingredients from BBLR before applying them to the asymmetric problem.
Throughout this subsection, references to equations and page numbers in
\cite{BBLR} refer to \href{https://arxiv.org/abs/1609.02539v1}{arXiv:1609.02539v1}.
For $\lambda=(\alpha,\beta,\gamma,\delta)$, define
\begin{equation}\label{eq:twisted-moment}
\begin{split}
I_\lambda(T;f,g)
=\int_{\mathbb R}\Phi(t/T)
&\zeta(s+\alpha)\zeta(s+\beta)
\zeta(\bar s+\gamma)\zeta(\bar s+\delta)\\
&\hspace{15mm}\cdot A_f(s)A_g(\bar s)\,\mathrm dt,
\qquad s=\tfrac12+it.
\end{split}
\end{equation}
Put
\[
\tau_{u,v}(n)=\sum_{de=n}d^{-u}e^{-v},
\qquad \lambda^*=(\gamma,\delta,\alpha,\beta),
\]
and, for coprime positive integers $h,k$, let
\begin{align*}
\mathcal A_\lambda
&=\frac{\zeta(1+\alpha+\gamma)\zeta(1+\alpha+\delta)
\zeta(1+\beta+\gamma)\zeta(1+\beta+\delta)}
{\zeta(2+\alpha+\beta+\gamma+\delta)},\\
\mathcal B_\lambda(h)
&=\prod_{\ell^\nu\parallel h}
\frac{\displaystyle\sum_{r\geq0}
\tau_{\alpha,\beta}(\ell^r)
\tau_{\gamma,\delta}(\ell^{r+\nu})\ell^{-r}}
{\displaystyle\sum_{r\geq0}
\tau_{\alpha,\beta}(\ell^r)
\tau_{\gamma,\delta}(\ell^r)\ell^{-r}},\\
Z_\lambda(h,k)
&=\mathcal A_\lambda\mathcal B_\lambda(h)
\mathcal B_{\lambda^*}(k).
\end{align*}
Here and below, $\ell$ denotes a prime when it occurs in an Euler product.
Writing $d=(h,k)$, define
\begin{equation}\label{eq:twisted-coefficient}
\mathcal M_\lambda(T;f,g)
=\sum_{h,k\geq1}
f\!\left(\frac{\log h}{L}\right)
g\!\left(\frac{\log k}{L}\right)
\frac{d}{hk}Z_\lambda(h/d,k/d).
\end{equation}
This is a finite coefficient sum and does not depend on $t$.
The six shift configurations and their phase exponents are
\begin{equation}\label{eq:twisted-six-swaps}
\begin{array}{c|c|c}
j&\lambda^{(j)}&e_j(\lambda)\\ \hline
0&(\alpha,\beta,\gamma,\delta)&0\\
1&(-\gamma,-\delta,-\alpha,-\beta)&-\alpha-\beta-\gamma-\delta\\
2&(-\gamma,\beta,-\alpha,\delta)&-\alpha-\gamma\\
3&(-\delta,\beta,\gamma,-\alpha)&-\alpha-\delta\\
4&(\alpha,-\gamma,-\beta,\delta)&-\beta-\gamma\\
5&(\alpha,-\delta,\gamma,-\beta)&-\beta-\delta.
\end{array}
\end{equation}
The same table defines $\sigma^{(j)}$ and
$e_j(\sigma)$ for scaled shifts:
\[
(\sigma/L)^{(j)}
=\sigma^{(j)}/L,
\qquad
e_j(\sigma/L)=e_j(\sigma)/L.
\]
We write
\begin{equation}\label{eq:twisted-main}
\mathfrak M_\lambda(T;f,g)
=\int_{\mathbb R}\Phi(t/T)
\sum_{j=0}^{5}
\left(\frac{t}{2\pi}\right)^{e_j(\lambda)}
\mathcal M_{\lambda^{(j)}}(T;f,g)\,\mathrm dt.
\end{equation}
Although individual summands may have poles when shifts coincide, their
sum has a removable continuation. This continuation, justified below,
is the meaning of \eqref{eq:twisted-main} at such configurations.
\begin{lemma}[Asymmetric twisted fourth moment]\label{thm:twisted}
Let $00$ the error is
\begin{equation}\label{eq:twisted-error}
O_{K,f,g,\Phi,\varepsilon}\!\left(
T^{1/2+a+b+\varepsilon}
+T^{3/4+(3a+5b)/8+\varepsilon}\right).
\end{equation}
The same statements hold with $f$ and $g$ interchanged in both the moment
and the main term. Thus the shorter polynomial may occur on either side
of the critical line.
\end{lemma}
\medskip
\noindent\textbf{Part I. Ingredients from BBLR.}
BBLR already allows two independent coefficient sequences.
We use that freedom without modification. The additional work will be
to apply the intermediate estimates in their proof without replacing
the two polynomial lengths by a common upper bound.
For the moment, suppose that the four signed scaled shifts
\begin{equation}\label{eq:twisted-separated-shifts}
-\sigma_1,\quad -\sigma_2,\quad \sigma_3,\quad \sigma_4
\end{equation}
are pairwise distinct. All preliminary estimates are uniform on compact
sets satisfying this condition. The restriction will be removed in
\hyperref[part:twisted-proof]{Part~II}.
Write $\lambda^\dagger=(-\gamma,-\delta,-\alpha,-\beta)$.
We choose an even entire function $G$, with $G(0)=1$, which decays rapidly
in vertical strips and vanishes at
\[
w=\pm\frac{\alpha+\gamma}{2},\quad
\pm\frac{\alpha+\delta}{2},\quad
\pm\frac{\beta+\gamma}{2},\quad
\pm\frac{\beta+\delta}{2}.
\]
One possible choice on the separated shift set is
\begin{equation}\label{eq:twisted-G}
G(w)=e^{w^2}
\prod_{u\in\{\alpha,\beta\}}
\prod_{v\in\{\gamma,\delta\}}
\left(1-\frac{4w^2}{(u+v)^2}\right).
\end{equation}
This choice is unchanged by permutations within either shift pair and by
$\lambda\mapsto\lambda^\dagger$.
Its coefficients grow at most as a fixed power of $L$ on compact
separated scaled shift sets. Such logarithmic factors will be absorbed
into $T^\varepsilon$; no uniform choice of $G$ at a collision is needed.
Define
\begin{align*}
\mathcal G_\lambda(w,t)
&=\prod_{u\in\{\alpha,\beta\}}
\frac{\Gamma((s+u+w)/2)}{\Gamma((s+u)/2)}
\prod_{v\in\{\gamma,\delta\}}
\frac{\Gamma((\bar s+v+w)/2)}{\Gamma((\bar s+v)/2)},\\
V_\lambda(x,t)
&=\frac{1}{2\pi i}\int_{(1)}
\frac{G(w)}{w}\mathcal G_\lambda(w,t)\pi^{-2w}x^{-w}
\,\mathrm dw,\\
\mathcal X_\lambda(t)
&=\chi(s+\alpha)\chi(s+\beta)
\chi(\bar s+\gamma)\chi(\bar s+\delta).
\end{align*}
Here $(c)$ denotes the upward-oriented line $\Re w=c$.
The approximate functional equation
\cite[Lemma~2.1 and equation~(3)]{BBLR} is
\begin{equation}\label{eq:twisted-afe}
\begin{split}
&\zeta(s+\alpha)\zeta(s+\beta)
\zeta(\bar s+\gamma)\zeta(\bar s+\delta)\\
&\quad=
\sum_{m,n\geq1}
\frac{\tau_{\alpha,\beta}(m)\tau_{\gamma,\delta}(n)}{\sqrt{mn}}
\left(\frac mn\right)^{-it}V_\lambda(mn,t)\\
&\qquad\quad+
\mathcal X_\lambda(t)
\sum_{m,n\geq1}
\frac{\tau_{-\gamma,-\delta}(m)\tau_{-\alpha,-\beta}(n)}{\sqrt{mn}}
\left(\frac mn\right)^{-it}V_{\lambda^\dagger}(mn,t)
+O_N(T^{-N})
\end{split}
\end{equation}
for every fixed $N>0$, uniformly for $T\leq t\leq2T$.
We call the two sums the direct and dual sums.
The estimates needed for these weights follow from Stirling's formula;
see \cite[equations~(4)--(6)]{BBLR}. In particular,
\begin{equation}\label{eq:twisted-gamma-estimates}
\mathcal X_\lambda(t)
=\left(\frac{t}{2\pi}\right)^{-\alpha-\beta-\gamma-\delta}
\bigl(1+O(T^{-1})\bigr),
\end{equation}
and, in a fixed vertical strip with $|\Im w|\leq\log T$,
\[
\pi^{-2w}\mathcal G_\lambda(w,t)
=\left(\frac{t}{2\pi}\right)^{2w}
\left(1+O\!\left(\frac{(1+|w|)^C}{T}\right)\right).
\]
The Gaussian in \eqref{eq:twisted-G} controls the complementary part of
the contour. For any fixed $B,i,j\geq0$, these estimates also give
\begin{equation}\label{eq:twisted-weight-decay}
x^i t^j\frac{\partial^{i+j}}{\partial x^i\partial t^j}
V_\lambda(x,t)
\ll_{B,i,j} L^C(1+x/T^2)^{-B}.
\end{equation}
The same bounds apply to the dual weight.
It is convenient to introduce the simpler Mellin weight
\begin{equation}\label{eq:twisted-Vstar}
V^*(x)=\frac{1}{2\pi i}\int_{(1)}
\frac{G(w)}{w}(2\pi)^{-2w}x^{-w}\,\mathrm dw.
\end{equation}
Its dependence on the shifts is suppressed in the notation.
By the preceding Stirling estimate, for every $\varepsilon>0$,
\begin{equation}\label{eq:twisted-weight-replacement}
V_\lambda(x,t)=V^*(x/t^2)+O(T^{-1+\varepsilon})
\qquad (x\geq1,
\ T\leq t\leq2T).
\end{equation}
The same replacement holds for $V_{\lambda^\dagger}$, with the same
function $V^*$.
The off-diagonal equation obtained from
\eqref{eq:twisted-afe} is
\[
hm_1m_2-kn_1n_2=\pm r,\qquad r>0.
\]
We record the estimate for a dyadic block in the form that retains its
dependence on the two amplifier lengths.
Let
\[
h\asymp A,\quad k\asymp B,\quad
m_i\asymp M_i,\quad n_i\asymp N_i,\quad r\asymp H,
\]
and put
\[
M=M_1M_2,\qquad N=N_1N_2,\qquad P=ABMN.
\]
Let $u_h,v_k$ be bounded coefficient sequences supported on the indicated
intervals, and let $W_0,\ldots,W_4$ be smooth weights supported in $[1,2]$.
Consider
\begin{equation}\label{eq:twisted-QD-sum}
\begin{split}
S_\pm=\sum_{hm_1m_2-kn_1n_2=\pm r}
u_hv_k\,
&W_0(r/H)W_1(m_1/M_1)W_2(m_2/M_2)\\
&\hspace{12mm}\cdot W_3(n_1/N_1)W_4(n_2/N_2).
\end{split}
\end{equation}
Small-power losses in the coefficients and in finitely many derivatives
of the weights are allowed and are included in $T^\varepsilon$ below.
All parameters are assumed to be bounded by fixed powers of $T$.
Suppose, for a sufficiently small auxiliary $\eta>0$, that
\begin{equation}\label{eq:twisted-QD-hypotheses}
\begin{gathered}
AM\asymp BN,\qquad MN\gg(AB)^{\delta_0},\qquad
M_1\ll M_2P^\eta,\qquad N_1\ll N_2P^\eta,\\
BN_1\leq AM_1,\qquad H\ll(AB)^{1/2+\eta},
\end{gathered}
\end{equation}
where $\delta_0>0$ is fixed.
The zero-frequency term in \cite[Proposition~3.1]{BBLR} is
\begin{equation}\label{eq:twisted-zero-frequency}
\begin{split}
\mathcal Z=\sum_{\substack{h,k,m_1,n_1,d,q\geq1\\
(hm_1,kn_1)=d}}
&u_hv_k W_0(dq/H)W_1(m_1/M_1)W_3(n_1/N_1)\\
&\quad\cdot\int_0^\infty
W_2\!\left(\frac{kn_1x}{dM_2}\right)
W_4\!\left(\frac{hm_1x}{dN_2}\right)\,\mathrm dx.
\end{split}
\end{equation}
Thus the shift is $r=dq$. Notice that $\mathcal Z$ is the same for the
two signs in \eqref{eq:twisted-QD-sum}.
The proof of that proposition splits the nonzero frequencies according
to $d=(hm_1,kn_1)$. More precisely, the large-$d$ range is
\[
d>D_0,\qquad
D_0=\min\left\{
(AB)^{1/2}(AM)^{-100\eta},
\left(\frac{AM_1N_1H}{BM_2}\right)^{1/3}
\right\},
\]
and the small-$d$ range is its complement.
\hyperlink{cite.BBLR}{Equation~(15)} of BBLR is used in the first range and \hyperlink{cite.BBLR}{equation~(16)} in the
second. After integration and summation over $d$, but before the final
symmetrization in $A,B$, the displayed estimates on \hyperlink{cite.BBLR}{pages~16--17} give
\begin{align}
E_{\mathrm I}
&\ll_\varepsilon T^\varepsilon(PH^2)^{1/4}
\left(AB+H^{1/4}A^{3/8}B^{1/8}P^{1/8}\right),
\label{eq:twisted-large-gcd}\\
E_{\mathrm {II}}
&\ll_\varepsilon T^\varepsilon(PH^2)^{1/4}
\left(AB+(ABH)^{1/4}P^{1/8}\right).
\label{eq:twisted-small-gcd}
\end{align}
For a prescribed $\varepsilon$, the auxiliary $\eta$ is chosen
sufficiently small to absorb the small-power losses in these estimates.
The preliminary replacement of the shifted weight in the Poisson
calculation contributes $O(T^\varepsilon H^2)$; see
\cite[equation~(14) and the preceding calculation]{BBLR}.
Consequently,
\begin{equation}\label{eq:twisted-QD-error}
S_\pm=\mathcal Z+
O\!\left(E_{\mathrm I}+E_{\mathrm {II}}+T^\varepsilon H^2\right).
\end{equation}
These are intermediate estimates from the proof, not the symmetrized
error bound in the statement of \hyperlink{cite.BBLR}{Proposition~3.1}.
To use \eqref{eq:twisted-QD-hypotheses}, the two factors of each divisor
function must first be ordered smoothly. Following
\cite[Section~4, pp.~21--25]{BBLR}, choose $\omega$ with
\[
\omega(x)+\omega(1/x)=1
\]
and with rapid decay as $x\to\infty$.
Its Mellin transform is required to vanish at
$\pm(\alpha-\beta)/2$ and $\pm(\gamma-\delta)/2$.
For example, one may define $\omega$ by inverse Mellin transformation of
\[
\widehat\omega(w)=\frac{e^{w^2}}{w}
\left(1-\frac{4w^2}{(\alpha-\beta)^2}\right)
\left(1-\frac{4w^2}{(\gamma-\delta)^2}\right).
\]
The pole at zero has residue one, and the remaining numerator is even.
Changing $w$ to $-w$ and moving the contour across zero gives the required
identity for $\omega$.
The identity
\begin{equation}\label{eq:twisted-ordering}
\bigl(\omega(m_1/m_2)+\omega(m_2/m_1)\bigr)
\bigl(\omega(n_1/n_2)+\omega(n_2/n_1)\bigr)=1
\end{equation}
therefore produces four ordered contributions.
The Mellin zeros ensure that removing this ordering later introduces no
residues at the two shift differences.
The weights in the moment are not initially products of five separate
functions. After dyadic localization, Mellin inversion separates their
dependence on the variables. Repeated integration by parts in the
Mellin transforms gives an integrable bound in all imaginary directions.
One can therefore apply \eqref{eq:twisted-QD-error} to the separated
weights and then undo the Mellin inversions. This is the separation and
recombination argument of \cite[Section~4, equation~(21)]{BBLR}.
Only finitely many derivative bounds are needed; small powers arising
from the localization are included in $T^\varepsilon$.
We spell out the resulting main-term formula, since this identifies the
four crossed terms in our moment. Put
\[
c_\ell(x,y,z)
=\frac{(1-\ell^{-1-x})(1-\ell^{-1-y})}{1-\ell^{-2-z}},
\]
and define the finite Euler factors
\begin{align*}
\eta_{\lambda;h}(w)
&=\sum_{e\mid h} e^{-\gamma+\delta-w}
\prod_{\ell\mid h/e}
c_\ell(\alpha-\beta+\gamma-\delta+w,
\gamma-\delta,\alpha-\beta+\gamma-\delta),\\
\eta_{\lambda;h,k}(w)
&=\eta_{\lambda;h}(w)\eta_{\lambda^*;k}(w).
\end{align*}
This is $\eta_{\alpha,\beta,\gamma,\delta,h,k}(0,0,w)$ in
\cite[equation~(23)]{BBLR}; only this specialization is needed here.
For a smooth weight $F(x,y,r)$, write
\[
\widehat F_3(x,y,w)=\int_0^\infty F(x,y,r)r^{w-1}\,\mathrm dr
\]
and
\begin{equation}\label{eq:twisted-Ftilde}
\begin{split}
\widetilde F_\lambda(x,y;h,k,d)
=\frac{1}{2\pi i}\int_{(1+\epsilon_0)}
&\widehat F_3(x,y,w)\zeta(w)
\zeta(1+\alpha-\beta+\gamma-\delta+w)\\
&\hspace{8mm}\cdot d^{-w}\eta_{\lambda;h,k}(w)\,\mathrm dw,
\end{split}
\end{equation}
where $\epsilon_0>0$ is fixed and small.
The weighted formula applies with $u_h,v_k$ supported on dyadic
intervals $[A,2A]$, $[B,2B]$, with $F$ supported on $r\asymp H$, and
with the derivative bounds
\[
\partial_x^i\partial_y^j\partial_r^k F(x,y,r)
\ll T^\varepsilon(1+x)^{-i}(1+y)^{-j}(1+r)^{-k},
\]
\[
\mathcal W^{(j)}(x)
\ll_{j,J}T^\varepsilon(1+x)^{-j}(1+x/X^2)^{-J}
\]
for fixed derivative orders and every fixed $J>0$.
An overall size factor in $F$ is extracted before using these bounds.
The same formula applies after summing smooth dyadic partitions.
For these weights, consider
\[
\mathcal S_\lambda(F,\mathcal W)
=\sum_{hm_1m_2-kn_1n_2=r>0}
\frac{u_hv_k}{m_1^\alpha m_2^\beta n_1^\gamma n_2^\delta}
F(hm_1m_2,kn_1n_2,r)\mathcal W(m_1m_2n_1n_2).
\]
The zero-frequency calculation in
\cite[Theorem~4.1 and its proof]{BBLR} produces four main terms,
\begin{equation}\label{eq:twisted-four-Q}
\mathcal Q_{\alpha,\beta,\gamma,\delta}
+\mathcal Q_{\beta,\alpha,\gamma,\delta}
+\mathcal Q_{\alpha,\beta,\delta,\gamma}
+\mathcal Q_{\beta,\alpha,\delta,\gamma},
\end{equation}
where the dependence on $F,\mathcal W,u,v$ is suppressed and
\begin{equation}\label{eq:twisted-Q-main}
\begin{split}
\mathcal Q_\lambda
={}&\frac{\zeta(1+\alpha-\beta)\zeta(1+\gamma-\delta)}
{\zeta(2+\alpha-\beta+\gamma-\delta)}
\sum_{d\geq1}\sum_{(h_0,k_0)=1}
\frac{u_{dh_0}v_{dk_0}\,d}
{(dh_0)^{1-\beta}(dk_0)^{1-\delta}}\\
&\quad\cdot\int_0^\infty
\mathcal W\!\left(\frac{x^2}{d^2h_0k_0}\right)
\widetilde F_\lambda(x,x;h_0,k_0,d)
x^{-\beta-\delta}\,\mathrm dx.
\end{split}
\end{equation}
The formula for the negative sign is obtained from the same
zero-frequency expression \eqref{eq:twisted-zero-frequency}.
There are two distinct errors in this passage. The nonzero frequencies
are controlled by \eqref{eq:twisted-QD-error}, before the Mellin
inversions are undone. The contour shifts used to remove the ordering
weights contribute an additional error. With $X=(MN)^{1/2}$, the bound
on \hyperlink{cite.BBLR}{page~25} of \cite{BBLR} is
\begin{equation}\label{eq:twisted-recovery-error}
E_{\mathrm{rec}}
\ll_\varepsilon T^\varepsilon(ABX^2)^{1/4}
H\bigl(A^{1/2}+B^{1/2}\bigr).
\end{equation}
This bound is for an unnormalized block; a scalar normalization of its
weight multiplies the error by the same scalar.
We will check this error separately rather than include it implicitly
in the quadratic-divisor estimate.
The connection between \eqref{eq:twisted-Q-main} and the factors
$Z_\lambda$ is the local identity
\cite[equation~(9), pp.~11--12]{BBLR}.
In our notation it gives, for $(h,k)=1$,
\begin{equation}\label{eq:twisted-eta-identity}
h^\beta k^\delta\eta_{\lambda;h,k}(\beta+\delta)
=\mathcal B_{(\alpha,-\delta,\gamma,-\beta)}(h)
\mathcal B_{(\gamma,-\beta,\alpha,-\delta)}(k).
\end{equation}
It is enough to verify this at prime powers; the identity for general
$h,k$ follows by multiplicativity.
For later use, put
\begin{equation}\label{eq:twisted-crossed-shift}
C_\lambda(z)=(\alpha+z,-\delta-z,\gamma+z,-\beta-z).
\end{equation}
The finite factor $\eta_{\lambda;h,k}(w)$ depends on the shifts only
through their differences. Applying
\eqref{eq:twisted-eta-identity} with all four shifts increased by $z$
therefore gives
\begin{equation}\label{eq:twisted-shifted-eta}
h^{\beta+z}k^{\delta+z}
\eta_{\lambda;h,k}(\beta+\delta+2z)
=\mathcal B_{C_\lambda(z)}(h)
\mathcal B_{C_\lambda(z)^*}(k).
\end{equation}
This identity will identify the entire contour integrand, not just its
residue at zero.
\medskip
\phantomsection\label{part:twisted-proof}
\noindent\textbf{Part II. Proof of the asymmetric moment formula.}
\begin{proof}[Proof of Lemma~\ref{thm:twisted}]
We first work on a compact separated set as in
\eqref{eq:twisted-separated-shifts}.
Write $f_h=f(\log h/L)$ and $g_k=g(\log k/L)$.
All implied constants may depend on this compact set, on the profiles,
and on $\Phi$. The profiles and the weight remain fixed as $T\to\infty$.
Insert \eqref{eq:twisted-afe} into \eqref{eq:twisted-moment} and expand
the two divisor functions. The direct contribution has the form
\begin{equation}\label{eq:twisted-expanded-moment}
\sum_{h,k,m_1,m_2,n_1,n_2\geq1}
\frac{f_hg_k}
{\sqrt{hkm_1m_2n_1n_2}\,
m_1^\alpha m_2^\beta n_1^\gamma n_2^\delta}
\int_{\mathbb R}\Phi(t/T)
\left(\frac{hm_1m_2}{kn_1n_2}\right)^{-it}
V_\lambda(m_1m_2n_1n_2,t)\,\mathrm dt.
\end{equation}
The dual contribution is treated in the same way.
We separate the diagonal $hm_1m_2=kn_1n_2$ and first estimate the error
in the off-diagonal calculation.
Choose small positive parameters $u$ and $\rho$.
By \eqref{eq:twisted-weight-decay}, the part with
$m_1m_2n_1n_2>T^{2+u}$ is negligible.
Repeated integration by parts in $t$ makes the terms with
\[
T\left|\log\frac{hm_1m_2}{kn_1n_2}\right|>T^\rho
\]
negligible as well. Here and below, negligible means $O(T^{-N})$ for
any prescribed fixed $N$, after increasing the number of integrations
by parts or the decay exponent.
On a remaining dyadic block, put
\[
Q=AB,\qquad R=\max\{A,B\},\qquad X=(MN)^{1/2},\qquad
\Lambda=T^\rho.
\]
The two products are comparable, so $AM\asymp BN\asymp\sqrt Q\,X$.
If their difference has size $H$, the preceding restrictions become
\begin{equation}\label{eq:twisted-block-ranges}
X\ll T^{1+u/2},\qquad
H\ll\Lambda\frac{\sqrt Q\,X}{T}.
\end{equation}
These are the two restrictions that will convert the divisor-sum
estimates into a moment estimate.
The coefficient normalization on the block is
$P^{-1/2}=(\sqrt Q\,X)^{-1}$, and the time integral has length $O(T)$.
Consequently the conversion factor for an unnormalized block error is
\begin{equation}\label{eq:twisted-normalization}
\frac{T}{\sqrt Q\,X}.
\end{equation}
All ratios between the exact square-root factors and their dyadic sizes
are included in the smooth weight.
We also record why the elementary weight replacements do not cost a
new leading error. For fixed $h,m_1,m_2$ and the difference $r$, the
number of factorizations of $hm_1m_2\mp r$ as $kn_1n_2$ is
$O_\varepsilon(T^\varepsilon)$.
There are $O(AMH)=O(\sqrt Q\,XH)$ choices before this factorization.
Thus the absolute coefficient mass of a near-diagonal block in
\eqref{eq:twisted-expanded-moment} is $O(T^\varepsilon H)$.
The shift-dependent powers are bounded on the truncated ranges.
In particular, replacing the gamma weight by
\eqref{eq:twisted-weight-replacement} is harmless.
Likewise, on the ranges \eqref{eq:twisted-block-ranges},
\[
\left(1\pm\frac{r}{kn_1n_2}\right)^{-it}
=\exp\!\left(\mp\frac{itr}{kn_1n_2}\right)
+O(\Lambda^2/T).
\]
Multiplying these errors by the coefficient mass and integrating in
$t$ gives a total of $O(T^{(a+b)/2+\varepsilon})$, after choosing
$u,\rho$ sufficiently small and summing the dyadic blocks.
This is smaller than the first term in \eqref{eq:twisted-error}.
We now verify that the refined quadratic-divisor bounds apply.
Because the two profiles have compact support inside their open
intervals, there are $a_0,b_0>0$ such that every nonzero coefficient
satisfies
\[
T^{a_0}\leq h\leq T^a,\qquad
T^{b_0}\leq k\leq T^b.
\]
Thus $Q\gg T^{a_0+b_0}$.
Choosing $\rho+u/2$ sufficiently small in
\eqref{eq:twisted-block-ranges} gives $H\ll Q^{1/2+\eta}$.
Moreover, a nonempty off-diagonal block has $H\gg1$, and hence
\[
MN=X^2\gg\frac{T^{2-2\rho}}{Q}.
\]
Since $Q\ll T^{a+b}$ and $a+b<1/2$, this implies
$MN\gg Q^{\delta_0}$ for some fixed $\delta_0>0$.
The inequality $a+b<1/2$ follows directly from
$4(a+b)\leq3a+5b<2$.
Use \eqref{eq:twisted-ordering} to order the factors within each side.
The rapid decay of $\omega$ permits the restrictions
$M_1\ll M_2P^\eta$ and $N_1\ll N_2P^\eta$, at a negligible cost.
Interchanging the two sides when necessary gives $BN_1\leq AM_1$.
This interchange leaves $Q$ and $R$ unchanged.
All conditions in \eqref{eq:twisted-QD-hypotheses} are now satisfied.
The estimates \eqref{eq:twisted-large-gcd} and
\eqref{eq:twisted-small-gcd} can be combined without replacing $A$ and
$B$ by the larger one. Indeed,
\[
A^{3/8}B^{1/8}\leq Q^{1/8}R^{1/4},\qquad
Q^{1/4}\leq Q^{1/8}R^{1/4},\qquad P=QX^2.
\]
Consequently the discrepancy between the divisor sum and its
zero-frequency term is
\begin{equation}\label{eq:twisted-discrepancy}
E\ll_\varepsilon T^\varepsilon
\left(
H^{1/2}Q^{5/4}X^{1/2}
+H^{3/4}Q^{1/2}R^{1/4}X^{3/4}
+H^2\right).
\end{equation}
Multiplying by \eqref{eq:twisted-normalization} and using the bound for
$H$ in \eqref{eq:twisted-block-ranges} gives
\begin{equation}\label{eq:twisted-block-error}
E_{\mathrm{block}}\ll_\varepsilon T^\varepsilon
\left(
\Lambda^{1/2}T^{1/2}Q
+\Lambda^{3/4}T^{1/4}Q^{3/8}R^{1/4}X^{1/2}
+\Lambda^2\frac{\sqrt Q\,X}{T}\right).
\end{equation}
Since $Q\ll T^{a+b}$, $R\ll T^b$ and $X\ll T^{1+u/2}$, the three
terms have respective bounds
\[
T^{1/2+a+b+\rho/2+\varepsilon},\qquad
T^{3/4+(3a+5b)/8+3\rho/4+u/4+\varepsilon},\qquad
T^{(a+b)/2+2\rho+u/2+\varepsilon}.
\]
The third is smaller than the first once the auxiliary parameters are
sufficiently small.
The main-term recovery error \eqref{eq:twisted-recovery-error} must also
be included. With the same normalization, its contribution is at most
\begin{align*}
\frac{T}{\sqrt Q\,X}E_{\mathrm{rec}}
&\ll_\varepsilon
T^\varepsilon\Lambda Q^{1/4}X^{1/2}
\bigl(A^{1/2}+B^{1/2}\bigr)\\
&\ll_\varepsilon
T^{1/2+(a+3b)/4+\rho+u/4+\varepsilon}.
\end{align*}
Because $(a+3b)/4\leq a+b$, this is absorbed by the first term of
\eqref{eq:twisted-error}.
There are only logarithmically many dyadic blocks.
For any prescribed final $\varepsilon>0$, choose all auxiliary
small-power parameters sufficiently small, including those needed for
the finitely many weight derivatives in the Mellin separation.
Absorbing the dyadic summation and these losses gives
\begin{equation}\label{eq:twisted-off-diagonal-error}
E_{\mathrm{off}}\ll_\varepsilon
T^{1/2+a+b+\varepsilon}
+T^{3/4+(3a+5b)/8+\varepsilon}.
\end{equation}
This is the step at which the separate lengths produce the condition
$3a+5b<2$.
We next identify the main terms, retaining their full dependence on the
shifts. On the diagonal, write $h=dh_0$, $k=dk_0$, with
$(h_0,k_0)=1$. The equation $hm=kn$ then gives
$m=k_0r$, $n=h_0r$.
The Euler factors defining $Z$ imply
\begin{equation}\label{eq:twisted-diagonal-series}
\sum_{r\geq1}
\frac{\tau_{\alpha,\beta}(k_0r)
\tau_{\gamma,\delta}(h_0r)}{r^{1+2w}}
=(h_0k_0)^w Z_{\lambda+w\mathbf1}(h_0,k_0),
\qquad \mathbf1=(1,1,1,1),
\end{equation}
initially in a half-plane of absolute convergence.
For example, a prime dividing $h_0$ contributes the shifted local
factor in $\mathcal B_{\lambda+w\mathbf1}(h_0)$; primes dividing $k_0$
contribute the other factor, and all remaining primes contribute
$\mathcal A_{\lambda+w\mathbf1}$.
Replacing $V_\lambda$ by $V^*$ on the diagonal costs only
$O(T^\varepsilon)$. To see this, write the common product as $q=hm=kn$.
On the truncated ranges the absolute diagonal coefficient sum is
bounded by $\sum_{q\leq T^C}d_3(q)^2/q=O(T^\varepsilon)$,
where $d_3(q)$ denotes the number of ordered factorizations of $q$ into
three positive integers. The error
$O(T^{-1+\varepsilon})$ in the weight is then integrated over an interval
of length $O(T)$.
Insert \eqref{eq:twisted-Vstar} and use
\eqref{eq:twisted-diagonal-series}. The factor $(h_0k_0)^w$ cancels the
corresponding factor from the Mellin weight. Thus the direct diagonal
contribution is, up to the preceding error,
\begin{equation}\label{eq:twisted-diagonal-contour}
\int_{\mathbb R}\Phi(t/T)\frac{1}{2\pi i}\int_{(1)}
\frac{G(w)}{w}\left(\frac{t}{2\pi}\right)^{2w}
\mathcal M_{\lambda+w\mathbf1}(T;f,g)
\,\mathrm dw\,\mathrm dt.
\end{equation}
Move the inner contour to $\Re w=-1/8$.
The possible zeta poles at $w=-(\alpha+\gamma)/2$ and the other three
cross sums are canceled by $G$.
The reciprocal zeta factor is regular in this strip, since its argument
has real part greater than one for large $T$.
Only the pole at $w=0$ remains, and its residue is
$\mathcal M_\lambda(T;f,g)$.
For completeness, a coarse bound suffices for the new contour.
The local geometric series defining $\mathcal B$, together with
polynomial vertical growth of the zeta factors, give on $\Re w=-1/8$
\[
Z_{\lambda+w\mathbf1}(h_0,k_0)
\ll_\varepsilon
(1+|\Im w|)^C(h_0k_0)^{1/2+\varepsilon}.
\]
Indeed the shifted divisor coefficients at a prime power grow at most
like its $1/8+O(1/L)$ power, times a divisor factor; the displayed
$1/2+\varepsilon$ exponent is a convenient weaker bound.
Summing with the coefficient $1/(dh_0k_0)$ gives
\[
\mathcal M_{\lambda+w\mathbf1}(T;f,g)
\ll_\varepsilon(1+|\Im w|)^C T^{(a+b)/2+\varepsilon}.
\]
Since $(t/2\pi)^{2\Re w}\ll T^{-1/4}$ and $G$ has Gaussian decay,
the new contour contributes
$O(T^{3/4+(a+b)/2+\varepsilon})$.
The dual diagonal is identical, with $\lambda^\dagger$ in place of
$\lambda$ and with the factor in
\eqref{eq:twisted-gamma-estimates}.
Consequently the two diagonal main terms are
\[
\int_{\mathbb R}\Phi(t/T)
\left(
\mathcal M_\lambda(T;f,g)
+\left(\frac{t}{2\pi}\right)^{-\alpha-\beta-\gamma-\delta}
\mathcal M_{\lambda^\dagger}(T;f,g)
\right)\,\mathrm dt.
\]
Their contour errors are absorbed in
\eqref{eq:twisted-off-diagonal-error}, because
\[
\frac{a+b}{2}\leq\frac{3a+5b}{8}.
\]
We explain the contribution corresponding to the $(\beta,\delta)$
swap. The other three follow by permuting $\alpha,\beta$ and
$\gamma,\delta$.
After the elementary replacements already justified, apply
\eqref{eq:twisted-Q-main} separately to the positive and negative
off-diagonal shifts, with
\[
\mathcal W(y)=V^*(y/t^2),\qquad
F_t^\pm(x,y,r)
=\frac{\vartheta(x/y)}{\sqrt{xy}}e^{\mp itr/y}
\psi\!\left(\frac{r^2T^{2-2\rho}}{xy}\right).
\]
Here $\psi$ is smooth, equals one near zero, and vanishes outside a
fixed compact interval. It implements the localization
\eqref{eq:twisted-block-ranges}; changing the cutoff outside that range
is negligible after integration in $t$.
The fixed smooth function $\vartheta$ equals one near $1$ and is supported
where $x/y\asymp1$. It is one on the retained terms and ensures that
$tr/y\ll\Lambda$ throughout the support of the weight.
Thus, after extracting its dyadic size, each fixed number of derivatives
costs at most a fixed power of $\Lambda$, as required for the Mellin
separation above. Since $\vartheta(1)=1$, it does not change the
zero-frequency main term.
The two original signed sums need not be equal.
However, their zero-frequency expressions evaluate the weights at
$x=y$, so their weights there add to
\[
F_t^+(x,x,r)+F_t^-(x,x,r)
=\frac{2}{x}\cos(tr/x)
\psi\!\left(\frac{r^2T^{2-2\rho}}{x^2}\right).
\]
This is why a cosine occurs without any symmetry assumption on $f,g$.
Taking the Mellin transform in $r$ and putting $r=xv$ gives
\[
2x^{w-1}\int_0^\infty
\cos(tv)\psi(v^2T^{2-2\rho})v^{w-1}\,\mathrm dv.
\]
As in \cite[Section~2.4, p.~10]{BBLR}, the smooth cutoff can be removed
in this Mellin calculation. Its transform has residue one at zero;
shifting the cutoff contour, or repeatedly integrating the tail by
parts, makes the error arbitrarily small after insertion in the
remaining integrals. The reason is that the oscillation at the cutoff
scale is $T^\rho\to\infty$.
The resulting expression is
\begin{equation}\label{eq:twisted-cosine-Mellin}
2x^{w-1}\Gamma(w)\cos(\pi w/2)t^{-w},
\end{equation}
interpreted by meromorphic continuation from $0<\Re w<1$.
The remaining $x$-integral is evaluated directly from
\eqref{eq:twisted-Vstar}:
\begin{equation}\label{eq:twisted-x-Mellin}
\begin{split}
&\int_0^\infty
V^*\!\left(\frac{x^2}{d^2h_0k_0t^2}\right)
x^{w-\beta-\delta-1}\,\mathrm dx\\
&\qquad=
\left(\frac{td\sqrt{h_0k_0}}{2\pi}\right)^{w-\beta-\delta}
\frac{G((w-\beta-\delta)/2)}{w-\beta-\delta}.
\end{split}
\end{equation}
Use also
\[
2(2\pi)^{-w}\Gamma(w)\cos(\pi w/2)\zeta(w)=\zeta(1-w).
\]
Substitution in \eqref{eq:twisted-Q-main}, followed by
$w=\beta+\delta+2z$, gives the following original crossed term, before
integration against $\Phi(t/T)\,\mathrm dt$:
\begin{equation}\label{eq:twisted-original-crossed}
\left(\frac{t}{2\pi}\right)^{-\beta-\delta}
\frac{1}{2\pi i}\int_{(c)}\frac{G(z)}{z}
\mathcal R_\lambda(z)\,\mathrm dz,
\end{equation}
where $c>0$ is fixed and small. Moving the transformed contour to
$(c)$ crosses no pole, since the shift-dependent poles lie within
$O(L^{-1})$ of zero. The arithmetic factor is
\begin{equation}\label{eq:twisted-crossed-arithmetic}
\begin{split}
\mathcal R_\lambda(z)
={}&\frac{\zeta(1+\alpha-\beta)\zeta(1+\gamma-\delta)}
{\zeta(2+\alpha-\beta+\gamma-\delta)}
\zeta(1+\alpha+\gamma+2z)\zeta(1-\beta-\delta-2z)\\
&\quad\cdot
\sum_{d\geq1}\sum_{(h_0,k_0)=1}
\frac{f_{dh_0}g_{dk_0}}
{d\,h_0^{1-\beta-z}k_0^{1-\delta-z}}
\eta_{\lambda;h_0,k_0}(\beta+\delta+2z).
\end{split}
\end{equation}
The factors of $2$ in the cosine transform and the change of Mellin
variable are accounted for by the functional equation and by
$\mathrm dw=2\,\mathrm dz$.
In particular, there is no additional numerical factor in
\eqref{eq:twisted-original-crossed}.
Now apply \eqref{eq:twisted-shifted-eta}.
The four zeta factors in \eqref{eq:twisted-crossed-arithmetic}, together
with its denominator, are exactly $\mathcal A_{C_\lambda(z)}$.
The powers $h_0^{\beta+z}k_0^{\delta+z}$ convert the finite Euler factor
into the two $\mathcal B$ factors. Since
$1/(dh_0k_0)=d/(hk)$, we obtain
\begin{equation}\label{eq:twisted-crossed-identification}
\mathcal R_\lambda(z)=\mathcal M_{C_\lambda(z)}(T;f,g).
\end{equation}
This is the precise arithmetic identification needed for the contour
pairing.
The matching dual term uses
$R_\lambda=(-\delta,-\gamma,-\beta,-\alpha)$, a permutation of
$\lambda^\dagger$ within its two pairs.
Its phase factor, after \eqref{eq:twisted-gamma-estimates}, is again
$(t/2\pi)^{-\beta-\delta}$.
Moreover, $C_{R_\lambda}(z)$ is $C_\lambda(-z)$ up to permutations
within the two pairs. The factors $Z$ and the auxiliary function $G$
are invariant under these permutations.
Changing $z$ to $-z$ in the dual integral therefore combines the two
terms into
\begin{equation}\label{eq:twisted-contour-pair}
\left(\frac{t}{2\pi}\right)^{-\beta-\delta}
\frac{1}{2\pi i}
\left(\int_{(c)}-\int_{(-c)}\right)
\frac{G(z)}{z}\mathcal M_{C_\lambda(z)}(T;f,g)
\,\mathrm dz.
\end{equation}
The replacement of the dual gamma factor has a relative error
$O(T^{-1})$. On either fixed contour, the same local-factor bounds as
above give an integrated error $O(T^{(a+b)/2+\varepsilon})$, which is
already admissible.
Besides $z=0$, the only possible poles between these contours are
\[
z=-\frac{\alpha+\gamma}{2},\qquad
z=-\frac{\beta+\delta}{2}.
\]
They are canceled by $G$.
The horizontal integrals tend to zero by its Gaussian decay.
The residue theorem thus evaluates \eqref{eq:twisted-contour-pair} as
\[
\left(\frac{t}{2\pi}\right)^{-\beta-\delta}
\mathcal M_{(\alpha,-\delta,\gamma,-\beta)}(T;f,g).
\]
Permuting the shifts within the two pairs gives the remaining three
one-swap terms in \eqref{eq:twisted-six-swaps}.
Together with the two diagonal terms, these are exactly
$\mathfrak M_\lambda(T;f,g)$.
We have therefore proved \eqref{eq:twisted-error} uniformly on compact
separated scaled shift sets.
We finally justify the continuation used in the statement and remove
the separation assumption.
Put
\[
x=(-\alpha,-\beta,\gamma,\delta).
\]
The six summands are indexed by the six choices of two of these four
coordinates for the first shift pair.
Each individual summand has possible simple poles only on the
hyperplanes $x_i=x_j$.
At a generic point of such a hyperplane, pair the terms obtained by
exchanging $x_i$ and $x_j$ between the two pairs.
Their phase factors and all regular Euler factors agree on the
hyperplane, whereas the polar factors are
$\zeta(1+x_i-x_j)$ and $\zeta(1-x_i+x_j)$, with opposite residues.
Thus the residues cancel.
To include intersections of hyperplanes, multiply the sum by
$\prod_{i0$.
Then
\[
\frac34+\frac{3a+5b}{8}=1-\frac{\Delta}{8},
\qquad
\frac12+a+b
=1-\frac{\Delta+b-a}{4}
\leq1-\frac{\Delta}{4}.
\]
Taking, for example, $0<\varepsilon<\Delta/16$ proves
\eqref{eq:twisted-asymptotic} uniformly on $K$.
Finally, the reality of $f,g,\Phi$ gives
\[
I_\lambda(T;g,f)
=\overline{I_{\overline{\lambda^*}}(T;f,g)}.
\]
The complete main term has the same symmetry, after interchanging
$h,k$ in its coefficient sums.
This proves the assertion with the profiles reversed.
\end{proof}
In particular, the formula applies locally uniformly to
$I_{(0,0,z/L,0)}(T;g,f)$, which is the orientation needed for the
$U$-numerator. Cauchy's integral formula also permits any fixed number
of derivatives with respect to the scaled shifts. The next subsection
uses this uniformity to combine the six main terms before specializing
the shifts.
\subsubsection{The $U$-numerator: from six swaps to the simplex kernel}
\label{sec:num-u}
To obtain the corresponding limits for the polynomial weights
$p$, we first consider the generating moment
\[
U_T(z)
=\int_{\R}\Phi(t/T)\zeta(s)^3\overline{\chi(s)}
A_f(\bar s)A_g(s)e^{z/2}
\zeta(\bar s+z/L)\,\dd t.
\]
The polynomial weights will be recovered by applying
$p(\partial_z)$ at $z=0$.
Our first step is to express $U_T(z)$ in terms of
the asymmetric twisted fourth moment established in
Section~\ref{sec:twisted}.
By the functional equation, the product of zeta
factors in $U_T(z)$ becomes
\[
\zeta(s)^2\zeta(\bar s)\zeta(\bar s+z/L),
\]
with two factors on each side of the critical line.
Thus, in the notation of the twisted fourth moment,
\begin{equation}\label{eq:num-u-fourth}
U_T(z)
=e^{z/2}I_{(0,0,z/L,0)}(T;g,f).
\end{equation}
Notice that the profiles appear in the order $(g,f)$,
since $A_g(s)$ and $A_f(\bar s)$ occur on opposite sides.
The asymmetric fourth-moment formula allows precisely
this ordering.
We can therefore apply that formula to $U_T(z)$.
It remains to extract the leading $L^9$ contribution from
its six main terms and show that their sum gives the common
kernel $K_z(a,b)$ introduced in Section~\ref{sec:num-kernel}.
The resulting locally uniform limit in $z$ will then allow
us to recover the $U$-numerator formula by differentiation
at $z=0$.
We first work with separated scaled shifts
$\sigma=(\sigma_1,\ldots,\sigma_4)$, so that the actual shifts
are given by $\lambda=\sigma/L$. This avoids the vanishing
denominators that may occur in the individual main terms
when shift parameters coincide. Although these terms may
appear singular, their sum will admit a regular continuation
to the shift configuration required for the $U$-numerator.
For convenience, we introduce
\[
D(\sigma)
=\prod_{i=1}^2\prod_{j=3}^4(\sigma_i+\sigma_j)
\]
and
\[
E_{r,s}(x)
=
\begin{cases}
\displaystyle
\frac{e^{-rx}-e^{-sx}}{s-r}, & r\neq s,\\[6pt]
xe^{-rx}, & r=s.
\end{cases}
\]
The four factors in $D(\sigma)$ arise from the polar
zeta factors associated with the two pairs of shifts.
The function $E_{r,s}$ will appear naturally when we
invert the remaining factors involving the amplifier
profiles.
Let $\sigma^{(j)}$ and $e_j(\sigma)$, for $0\leq j\leq5$,
denote the transformed shifts and the corresponding
exponents from \eqref{eq:twisted-six-swaps}. Each of
these six configurations contributes one main term.
To identify their leading contributions, we use the
polar approximation
\[
\frac{1}{L}\zeta\left(1+\frac{w}{L}\right)
\longrightarrow \frac{1}{w},
\qquad L\to\infty,
\]
for fixed $w\neq0$.
For the unswapped term, the resulting polar factors
have the form
\[
\frac{1}{
D(\sigma)(u+v)
(u+\sigma_3)(u+\sigma_4)
(v+\sigma_1)(v+\sigma_2)
}.
\]
There are nine denominator factors in total, which
accounts for the leading power $L^9$.
The variables $u$ and $v$ arise from Mellin inversion
of the two amplifier profiles. For a profile $h$, we use
the convention
\[
\widetilde h(w)=\int_0^\infty h(x)e^{wx}\,\dd x,
\qquad
h(\log n/L)
=\frac{1}{2\pi i}\int_{(c)}
\widetilde h(w)n^{-w/L}\,\dd w.
\]
Here $u$ and $v$ correspond to the profiles $g$ and
$f$, respectively. The factors involving
these variables can be inverted using
\[
\frac{1}{u+v}
=\int_0^\infty e^{-t(u+v)}\,\dd t
\]
and
\[
\mathcal L^{-1}
\left(\frac{1}{(u+r)(u+s)}\right)(x)
=E_{r,s}(x).
\]
The first identity introduces a common integration
variable $t$, while the second produces the two
functions $E_{r,s}$ associated with the profiles.
Let \(A\) and \(B\) denote the logarithmic positions associated with the profiles \(f\) and \(g\), respectively. Since we apply the twisted fourth-moment formula to \(I_\lambda(T;g,f)\), the first amplifier has position \(B\) and the second has position \(A\). Applying these identities
to each of the six shifted main terms gives the
following leading kernel:
\begin{equation}\label{eq:num-six-kernel}
\begin{aligned}
\int_0^{\min(A,B)}
\sum_{j=0}^5
&\frac{e^{e_j(\sigma)}}{D(\sigma^{(j)})}\\
&\times
E_{\sigma^{(j)}_3,\sigma^{(j)}_4}(B-t)
E_{\sigma^{(j)}_1,\sigma^{(j)}_2}(A-t)
\,\dd t.
\end{aligned}
\end{equation}
Here the exponential factor $e^{e_j(\sigma)}$
arises from the corresponding power
$(\tau/2\pi)^{e_j(\sigma)/L}$ in the twisted
fourth-moment formula. Since $\tau\asymp T$
on the support of the smoothing function,
this factor converges to $e^{e_j(\sigma)}$.
The expression in \eqref{eq:num-six-kernel} gives the leading
contribution for fixed logarithmic positions $A$ and $B$.
To recover the main term of the twisted fourth moment,
we integrate this kernel against the two amplifier profiles,
as follows from their Mellin inversion.
The integration against the smoothing function contributes
$T\widehat\Phi(0)$, while the regular part of the
Euler product gives the arithmetic constant $a_3$.
Together with the previously obtained factor $L^9$
and the six swap contributions, this yields
$a_3TL^9\widehat\Phi(0)$ times the profile average
of \eqref{eq:num-six-kernel}, up to an error of
$o(TL^9)$.
It remains to combine the six summands into a
determinant representation, which will allow us
to pass to the required coincident shifts and
recover the simplex kernel.
To justify the limiting Mellin calculation, including its tails, fix a
compact set $K$ of scaled shifts such that
$D(\sigma^{(j)})\neq 0$ for $0\leq j\leq 5$. Choose $R>0$ bounding all components of the six transformed
shift configurations on $K$, and take $c>R+1$.
On the vertical contours
$u=c+i\xi$, $v=c+i\upsilon$, the five arguments involving $u,v$
in the unswapped term are
\[
\mathcal W=\{u+v,u+\sigma_3,u+\sigma_4,v+\sigma_1,v+\sigma_2\}.
\]
For $w\in\mathcal W$, we have $\Re w>1$, and absolute convergence gives,
for $L\geq1$,
\[
|L^{-1}\zeta(1+w/L)|
\leq L^{-1}\zeta(1+\Re w/L)
\leq L^{-1}+(\Re w)^{-1}\leq2.
\]
The same bound applies to all six swap configurations.
The four remaining zeta factors depend only on the shifts. Since
$D(\sigma^{(j)})\neq 0$ on the compact set $K$, their scaled
arguments stay uniformly away from zero, so these factors are
uniformly bounded after division by $L$.
As in the proof of Lemma~\ref{lem:gram-diagonal}, cancellation of the first-order
terms in the local Euler factors yields
$1+O(\ell^{-2+\delta})$ for some fixed $0<\delta<1$.
This estimate is uniform on the Mellin contours for sufficiently
large $T$, since the real parts of the scaled shifts are
$O_K(L^{-1})$ and $|\ell^{-w}|=\ell^{-\Re w}$.
Consequently, the regular Euler product is uniformly bounded. Since $f,g\in C_c^\infty$, their
Mellin transforms satisfy
\[
|\widetilde g(c+i\xi)|\ll_{g,c}(1+|\xi|)^{-2},
\qquad
|\widetilde f(c+i\upsilon)|\ll_{f,c}(1+|\upsilon|)^{-2}.
\]
Combining these estimates with the preceding bounds for the five
Mellin-dependent zeta factors, the normalized integrand is bounded
in absolute value by
\begin{equation}\label{eq:num-mellin-majorant}
\frac{C_{K,f,g}}{(1+|\xi|)^2(1+|\upsilon|)^2},
\end{equation}
uniformly for $\sigma\in K$ and all sufficiently large $T$.
The right-hand side is integrable over $\mathbb R^2$.
On bounded segments of the Mellin contours, the nine normalized
zeta factors converge uniformly to the reciprocals of their
arguments. The regular Euler factor is holomorphic near the
origin and takes the value $a_3$ there, as established by the
nine-edge Euler-product calculation in Lemma~\ref{lem:gram-diagonal}. Since all its
scaled arguments tend uniformly to zero, it also converges
uniformly to $a_3$.
The majorant in \eqref{eq:num-mellin-majorant} controls the
remaining tails uniformly for $\sigma\in K$. Together with
the uniform convergence on bounded contour segments, this
gives uniform convergence of the normalized Mellin integrals
on $K$. Restoring the main-term prefactor, the resulting
error is $o(TL^9)$.
We now combine the six summands in \eqref{eq:num-six-kernel} into a determinant.
Put
\[
x=(-\sigma_1,-\sigma_2,\sigma_3,\sigma_4),\qquad
\rho(u,v)=(-1,v-1,-u,0),
\]
and define the Vandermonde determinant
\[
\Delta(x)=\prod_{0\leq i0$ with $a+b<1$, we have
\begin{equation}\label{eq:num-reflection}
e^{-z/2}J(a,b;-z)=K_z(a,1-b)
\qquad (z\in\C).
\end{equation}
\end{lemma}
\begin{proof}
Both sides are entire in $z$, so it suffices to establish
the identity for $z\ne0$.
Write
\[
E(a,z)=e^{az}-1-az
\]
and define
\begin{align*}
N(a,b,z)={}&z(1-z/2)E(a,z)E(b,z)\\
&-\frac{z^3}{2}
\bigl(a^2E(b,z)+b^2E(a,z)\bigr)\\
&+\frac{a^2b^2z^4}{4}(e^z-1).
\end{align*}
We first evaluate $J(a,b;-z)$. From the definitions
of $B_a,C_a$ and $H_m$ in the preceding subsection,
\[
B_a(-z)=\frac{E(a,z)}{z^2},
\qquad
C_a(-z)=-\frac{E(a,z)-a^2z^2/2}{z^3},
\]
and
\[
H_m(-z)=
\frac{e^z-\sum_{j=0}^m z^j/j!}{z^{m+1}},
\qquad m=0,1,2.
\]
Substituting these expressions into the definition of $J$
and collecting terms gives
\begin{equation}\label{eq:num-reflection-j}
z^7J(a,b;-z)=N(a,b,z).
\end{equation}
We next evaluate $K_z(a,1-b)$. Since $a+b<1$,
we have $\min(a,1-b)=a$. For $0\le t\le a$,
the four translated simplex nodes are
\[
0,\qquad a-t,\qquad b+t,\qquad 1.
\]
By the divided-difference formula for the simplex
expectation in the definition of $K_z$, we obtain
\begin{align*}
2z^3e^{z/2}K_z(a,1-b)
=\int_0^a\bigl[
&-(b-a+2t)(1-a+t)(1-b-t)\\
&+e^{z(a-t)}(b+t)(1-b-t)\\
&-e^{z(b+t)}(a-t)(1-a+t)\\
&+e^z(a-t)(b+t)(b-a+2t)
\bigr]\,\mathrm{d}t.
\end{align*}
This expression remains valid when two simplex nodes
coincide, by continuity.
Integrating the polynomial and exponential terms gives
\[
2z^3e^{z/2}K_z(a,1-b)
=\frac{2N(a,b,z)}{z^4}.
\]
Consequently,
\begin{equation}\label{eq:num-reflection-k}
e^{z/2}K_z(a,1-b)=\frac{N(a,b,z)}{z^7}.
\end{equation}
Comparing \eqref{eq:num-reflection-j} and
\eqref{eq:num-reflection-k}, we conclude that
\[
e^{-z/2}J(a,b;-z)=K_z(a,1-b)
\]
for $z\ne0$. Since both sides are entire, the identity
also holds at $z=0$.
\end{proof}
We can now complete the proof of Proposition~\ref{prop:numerator}.
Combining the locally uniform limit
\eqref{eq:num-v-generator-limit} with
Lemma~\ref{lem:num-reflection} yields
\[
\frac{V_T(z)}{a_3TL^9}
\longrightarrow
\widehat\Phi(0)
\int_0^\eta\int_0^\eta
f(a)g(b)K_z(a,1-b)\,\mathrm{d}a\,\mathrm{d}b.
\]
The convergence is locally uniform near $z=0$.
Hence Cauchy's integral formula allows us to apply
$p(\partial_z)$ and then set $z=0$. By the definition
of $K_p$, this gives
\[
k_V(p;f,g)
=\int_0^\eta\int_0^\eta
f(a)g(b)K_p(a,1-b)\,\mathrm{d}a\,\mathrm{d}b.
\]
This proves the $V$-numerator formula. Together with
the $U$-numerator limit established in
Section~\ref{sec:num-u}, it completes the proof of
Proposition~\ref{prop:numerator}.
The reflection identity concerns the limiting kernels
only; the analytic $V$-moment estimate still requires
the short-length condition $\eta<1/4$.
\subsection{Denominator}
\label{sec:denominator}
Unlike the numerator, which only involves correlations of individual
tests with $\zeta(s)^3$, the denominator requires all pairwise products
of the projected tests, including interactions between the two families
$U$ and $V$. By the projection identity \eqref{eq:projection}, each Gram entry splits
into an ordinary and a reflected product. Using the functional equation,
these products reduce to two types of moments: balanced moments, with
two amplifier factors on each side, and unbalanced moments, with three
factors on one side and one on the other. The only exception is the
reflected $VV$ product, whose oscillatory contribution will be shown
to be negligible.
We first establish the necessary approximation for the functional
equation of centered derivatives, which introduces the polynomial
reflection $p(x)\mapsto p(-x)$. We then evaluate the balanced and
unbalanced moments separately. In both cases, we isolate the diagonal
contribution, control the off-diagonal terms, and apply the nine-edge
diagonal limit from Lemma~\ref{lem:gram-diagonal}. The balanced case uses a smoothed
approximate functional equation and produces an additional length
constraint, whereas the unbalanced case uses a one-sided expansion.
This leads to two explicit kernels, $B_{p,q}$ and $M_{p,q}$,
from which we obtain all entries of the Gram matrix.
Throughout this section, the real profiles $f,g,h,k$ are continuous
and supported in $[0,A_i]$, for $i=1,2,3,4$, respectively, with
\[
S_*:=A_1+A_2+A_3+A_4<1.
\]
This condition holds for every pair of admissible tests from
Section~\ref{sec:setup} and ensures sufficient control of the off-diagonal terms.
The variables $a,b,c,d$ below are integration variables, not support
bounds. The polynomials $p,q$ have degree at most four.
\subsubsection{The functional-equation factor}
\begin{lemma}\label{lem:reflection-factor}
For every fixed $R>0$ and integer $j\geq 0$, we have
\[
\partial_z^j\bigl(R_T(t,z)-1\bigr)
\ll_{R,j} L^{-1}
\]
uniformly for $T\leq t\leq 2T$ and $|z|\leq R$.
\end{lemma}
\begin{proof}
By Stirling's formula, the logarithmic derivative
of $\chi$ satisfies
\[
\frac{\chi'}{\chi}(s+w)
=-\log\frac{t}{2\pi}+O(T^{-1})
\]
uniformly for $T\leq t\leq 2T$ and $|w|\leq R/L$.
Consequently,
\begin{align*}
\log R_T(t,z)
&=z+\int_0^{z/L}\frac{\chi'}{\chi}(s+w)\,\mathrm{d}w\\
&=z-\frac{z}{L}\log\frac{t}{2\pi}
+O_R\left(\frac{1}{TL}\right)\\
&=-\frac{z}{L}\log\frac{t}{2\pi T}
+O_R\left(\frac{1}{TL}\right).
\end{align*}
Since $t/T\in[1,2]$, the logarithm in the last
expression is bounded. Thus
\[
R_T(t,z)=1+O_R(L^{-1})
\]
uniformly on every fixed disc. The bounds for
the derivatives follow from Cauchy's integral
formula, applied on a slightly larger fixed disc.
\end{proof}
Replacing $R_T(t,z)$ by $1$ in the centered
functional equation formally corresponds to the
polynomial reflection $p\mapsto p^-$.
This replacement is not an exact identity between
centered derivatives. In the moment calculations
below, we justify it by showing that the resulting
error is negligible after integration.
\subsubsection{The balanced kernel}
For pairs of tests from the same family, the terms \(H\overline J\) in \eqref{eq:projection} take the form
\[
\begin{aligned}
V_{p;f,g}\overline{V_{q;h,k}}
&=C_p(s)C_q(\bar s)\,
A_f(s)A_g(s)A_h(\bar s)A_k(\bar s),\\[4pt]
U_{p;f,g}\overline{U_{q;h,k}}
&=C_p(s)C_q(\bar s)\,
A_f(s)A_k(s)A_g(\bar s)A_h(\bar s).
\end{aligned}
\]
Thus both are instances of the same balanced moment, up to a relabeling of the profiles.
For $a,b,c,d\geq0$ with $a+b+c+d<1$, define
\[
\mathcal R(a,b;c,d)=
\left\{
r\in\mathbb R_{\geq0}^{2\times2}:
\begin{aligned}
r_{11}+r_{12}&\leq a, &
r_{21}+r_{22}&\leq b,\\
r_{11}+r_{21}&\leq c, &
r_{12}+r_{22}&\leq d
\end{aligned}
\right\}.
\]
Put
\[
t_r=\sum_{i,j}r_{ij},\qquad
\Delta=\frac{a+b-c-d}{2},\qquad
w_r=\frac{1-a-b-c-d+2t_r}{2},
\]
and define
\begin{equation}\label{eq:gram-balanced-kernel}
B_{p,q}(a,b;c,d)=
\int_{\mathcal R(a,b;c,d)}
\int_{-w_r}^{w_r}
p(\Delta-x)q(-\Delta-x)\,\dd x\,\dd r.
\end{equation}
\begin{proposition}[Balanced moment]\label{prop:gram-balanced}
The normalized integral
\[
\frac1{a_3TL^9}\int_{\R}\Phi(t/T)
C_p(s)C_q(\bar s)A_f(s)A_g(s)A_h(\bar s)A_k(\bar s)\,\dd t
\]
converges to
\[
\widehat\Phi(0)\int f(a)g(b)h(c)k(d)
B_{p,q}(a,b;c,d)\,\dd a\,\dd b\,\dd c\,\dd d.
\]
\end{proposition}
\begin{proof}
We begin by introducing two complex shifts in order to handle
the centered derivatives. For $z,w$ in fixed complex discs, put
\[
s_1=s+\frac{z}{L},
\qquad
s_2=\bar s+\frac{w}{L},
\]
and consider the shifted integral
\[
\begin{aligned}
I(z,w):=\int_{\mathbb R}
&\Phi(t/T)e^{(z+w)/2}\zeta(s_1)\zeta(s_2)\\
&\quad\cdot A_f(s)A_g(s)
A_h(\bar s)A_k(\bar s)\,\mathrm{d}t.
\end{aligned}
\]
By the definition of the centered derivatives in Section~\ref{sec:setup},
the integral in Proposition~\ref{prop:gram-balanced} is exactly
\[
\left.
p(\partial_z)q(\partial_w)I(z,w)
\right|_{z=w=0}.
\]
It therefore suffices to analyze $I(z,w)$ uniformly for
bounded $z,w$, and to apply the differential operators
at the end of the proof.
To evaluate $I(z,w)$, we use a smoothed approximate
functional equation for the product $\zeta(s_1)\zeta(s_2)$.
This expresses the product as the sum of two rapidly
convergent double Dirichlet series, one coming directly
from the Dirichlet series and the other from the
functional equation.
We first specify the smoothing. Write
\[
\Gamma_{\mathbb R}(v)
=\pi^{-v/2}\Gamma(v/2),
\]
so that
\[
\Gamma_{\mathbb R}(v)\zeta(v)
=\Gamma_{\mathbb R}(1-v)\zeta(1-v).
\]
For $\nu=1,2$, define
\[
s_\nu^+=s_\nu,
\qquad
s_\nu^-=1-s_\nu.
\]
We choose the even smoothing function
\[
\ G_{z,w}(u)
=e^{u^2}
\prod_{\nu=1}^{2}
\left(1-\frac{u^2}{s_\nu^2}\right)
\left(1-\frac{u^2}{(1-s_\nu)^2}\right).
\]
It satisfies $\ G_{z,w}(0)=1$. The polynomial
factors cancel the poles of the completed zeta functions
encountered when shifting the Mellin contour.
For $x>0$, define the two weights
\[
\ W^\pm_{z,w}(x;t)
=\frac{1}{2\pi i}\int_{(1)}
\ G_{z,w}(u)
\frac{
\Gamma_{\mathbb R}(s_1^\pm+u)
\Gamma_{\mathbb R}(s_2^\pm+u)
}{
\Gamma_{\mathbb R}(s_1^\pm)
\Gamma_{\mathbb R}(s_2^\pm)
}
x^{-u}\frac{\mathrm{d}u}{u}.
\]
Here $\int_{(1)}$ denotes integration along the
vertical line $\Re u=1$, oriented upwards.
The smoothed approximate functional equation now gives
\[
\begin{aligned}
\zeta(s_1)\zeta(s_2)
={}&
\sum_{m,n\ge1}
\frac{\ W^+_{z,w}(mn;t)}
{m^{s_1}n^{s_2}}\\
&+\chi(s_1)\chi(s_2)
\sum_{m,n\ge1}
\frac{\ W^-_{z,w}(mn;t)}
{m^{1-s_1}n^{1-s_2}}.
\end{aligned}
\]
With our choice of the pole-corrected smoothing
function, this is an exact identity. Indeed, it follows
by shifting the Mellin contour from $\Re u=1$ to
$\Re u=-1$, applying the functional equation of the
completed zeta function, and using the evenness of
$\ G_{z,w}$. Only the residue at $u=0$ remains.
We call the first sum the direct term and the
second sum the dual term.
For the subsequent moment calculation it is convenient
to interchange $m$ and $n$ in the dual term. Since
the weights depend on the summation variables only
through their product $mn$, this gives
\[
\begin{aligned}
\zeta(s_1)\zeta(s_2)
=\sum_{m,n\ge1}\frac{(m/n)^{-it}}{\sqrt{mn}}
\Bigl(
& m^{-z/L}n^{-w/L}
\ W^+_{z,w}(mn;t)\\
&+\chi(s_1)\chi(s_2)
m^{w/L}n^{z/L}
\ W^-_{z,w}(mn;t)
\Bigr).
\end{aligned}
\]
In this form, both sums have the same oscillatory
factor $(m/n)^{-it}$. Their difference lies in the
shift-dependent coefficients and in the factor
$\chi(s_1)\chi(s_2)$ appearing in the dual term.
We will insert this decomposition into $I(z,w)$,
expand the four Dirichlet polynomials, and show that
only the diagonal contributions survive in the
leading term. The dependence on $z,w$ will be kept
throughout, allowing us to recover $C_p$ and $C_q$
by differentiation at the end.
Fix $\delta>0$ such that
$S_*+\delta<1$. Write $h_1,h_2$ for the summation variables in
$A_f(s)A_g(s)$ and $k_1,k_2$ for those in
$A_h(\bar s)A_k(\bar s)$, and set
\[
H=h_1h_2,\qquad K=k_1k_2,\qquad
\mathcal A(h_1,h_2,k_1,k_2)
=f\!\left(\frac{\log h_1}{L}\right)
g\!\left(\frac{\log h_2}{L}\right)
h\!\left(\frac{\log k_1}{L}\right)
k\!\left(\frac{\log k_2}{L}\right).
\]
The support assumptions imply $HK\le T^{S_*}$. Inserting the
preceding decomposition and expanding the four Dirichlet polynomials,
we obtain terms with the common oscillatory factor
\[
\left(\frac{Hm}{Kn}\right)^{-it}.
\]
We first simplify the smoothing weights and show that the terms with
$Hm\ne Kn$ are negligible.
Define
\[
V_0(x)=\frac{1}{2\pi i}\int_{(1)}e^{u^2}x^{-u}\,\frac{\mathrm{d}u}{u}
=\frac{1}{\sqrt\pi}\int_{\frac12\log x}^{\infty}e^{-v^2}\,\mathrm{d}v.
\]
For fixed bounded shifts $z,w$, Stirling's formula gives, uniformly for
$T\le t\le2T$ and $u$ in a fixed vertical strip with
$|\Im u|\le\log T$,
\[
\frac{\Gamma_{\mathbb R}(s_1^\pm+u)\Gamma_{\mathbb R}(s_2^\pm+u)}
{\Gamma_{\mathbb R}(s_1^\pm)\Gamma_{\mathbb R}(s_2^\pm)}
=\left(\frac{t}{2\pi}\right)^u
\left(1+O\!\left(\frac{(1+|u|)^C}{T}\right)\right),
\]
where $C$ is an absolute fixed constant. Moreover, the polynomial
factor in $G_{z,w}(u)$ is
$1+O((1+|u|)^C/T^2)$ in this range. The
Gaussian makes the complementary part of either vertical integral
negligible. When the Mellin integrands defining $W_{z,w}^\pm$ and
$V_0(2\pi x/t)$ are subtracted, their numerator vanishes at $u=0$.
We may therefore move the contour of the difference to $\Re u=0$
without crossing a pole. The preceding estimates, including the
cancellation of the factor $1/u$ at the origin, yield
\begin{equation}\label{eq:gram-weight-replacement}
W_{z,w}^\pm(x;t)
=V_0\!\left(\frac{2\pi x}{t}\right)+O(T^{-1})
\end{equation}
uniformly for $x>0$, $T\le t\le2T$, and $z,w$ in fixed complex
discs. The same estimates give the corresponding bounds for any
fixed number of $t$-derivatives, with an additional factor $T^{-1}$
for each derivative.
For every fixed $B>1$, moving the defining Mellin contour to
$\Re u=B$ gives
\[
W_{z,w}^\pm(x;t),\quad V_0(2\pi x/t)
\ll_B (t/x)^B.
\]
Consequently, the portions of the two double sums with
$mn>T^{1+\delta}$ are negligible. Indeed, the total absolute
coefficient sum over the four amplifier variables is
$O(T^{S_*/2})$, while
\[
\sum_{mn>X}\frac{1}{\sqrt{mn}}\left(\frac{T}{mn}\right)^B
\ll_B T^B X^{1/2-B}\log(2X).
\]
The shift-dependent powers in the tails are at most
$(mn)^{O(1/L)}$, and are absorbed by increasing $B$ slightly.
Taking $X=T^{1+\delta}$ and then $B$ sufficiently large makes the
integrated tail $o(TL^9)$. On the truncated range, the total absolute
coefficient sum satisfies
\[
\sum_{\substack{h_1,h_2,k_1,k_2\\mn\le T^{1+\delta}}}
\frac{|\mathcal A(h_1,h_2,k_1,k_2)|}{\sqrt{mnHK}}
\ll T^{(1+S_*+\delta)/2}(\log T)^C.
\]
Thus the error $O(T^{-1})$ in \eqref{eq:gram-weight-replacement}, integrated over a time interval
of length $O(T)$, contributes
$O(T^{(1+S_*+\delta)/2}(\log T)^C)=o(TL^9)$.
All shift-dependent powers are bounded on the truncated range, and
$\chi(s_1)\chi(s_2)$ is uniformly bounded.
It follows that, up to $o(TL^9)$, the shifted integral is the sum over
$h_1,h_2,k_1,k_2$ and $mn\le T^{1+\delta}$ of
\[
\begin{aligned}
&\frac{\mathcal A(h_1,h_2,k_1,k_2)}{\sqrt{mnHK}}
\int_{\mathbb R}\!\Phi(t/T)
\left(\frac{Hm}{Kn}\right)^{-it}
V_0\!\left(\frac{2\pi mn}{t}\right)e^{(z+w)/2}\\
&\hspace{45mm}\cdot
\left(m^{-z/L}n^{-w/L}
+\chi(s_1)\chi(s_2)m^{w/L}n^{z/L}\right)\mathrm{d}t.
\end{aligned}
\]
The amplitude after removal of the displayed oscillatory factor has
$j$-th $t$-derivative $O_j(T^{-j})$, uniformly in all the truncated
indices and in the fixed shift discs. If $Hm\ne Kn$ and the ratio
$Hm/(Kn)$ is close to $1$, integrality gives
\[
\left|\log\frac{Hm}{Kn}\right|
\gg \frac{1}{\max(Hm,Kn)}
\gg (mnHK)^{-1/2}.
\]
Away from $1$ the logarithm is bounded below by a positive constant.
Since $mnHK\le T^{1+\delta+S_*}$, in either case
\[
T\left|\log\frac{Hm}{Kn}\right|
\gg T^{(1-S_*-\delta)/2}.
\]
Repeated integration by parts therefore bounds each off-diagonal
time integral by
\[
\ll_N T\left(1+T\left|\log\frac{Hm}{Kn}\right|\right)^{-N}.
\]
Combining this with the preceding absolute coefficient bound, and
choosing $N$ sufficiently large, shows that all terms with
$Hm\ne Kn$ contribute $o(TL^9)$.
We are left with the diagonal $Hm=Kn$. On this diagonal, Stirling's
formula for the functional-equation factor gives
\[
\chi(s_1)\chi(s_2)
=\exp\!\left(-\frac{z+w}{L}\log\frac{t}{2\pi}\right)
\bigl(1+O(T^{-1})\bigr)
=e^{-(z+w)}\bigl(1+O(L^{-1})\bigr)
\]
uniformly for $T\le t\le2T$. The nine-edge diagonal limit of
Lemma~\ref{lem:gram-diagonal} shows that the diagonal sum, restricted to
$mn\le T^{1+\delta}$ and normalized by $L^9$, is bounded.
Hence replacing $\chi(s_1)\chi(s_2)$ by $e^{-(z+w)}$ changes the
integral by $o(TL^9)$.
We next remove the remaining Gaussian smoothing. Put
\[
Q=\frac{\log(mn)}{L},\qquad t=T\tau,\qquad1\le\tau\le2.
\]
Then
\[
V_0\!\left(\frac{2\pi mn}{t}\right)
=\frac{1}{\sqrt\pi}
\int_{\frac12\{L(Q-1)+\log(2\pi/\tau)\}}^\infty
e^{-v^2}\,\mathrm{d}v.
\]
For every $\varepsilon>0$ this converges uniformly to $1$ on
$Q\le1-\varepsilon$, and uniformly to $0$ on
$Q\ge1+\varepsilon$, independently of $\tau\in[1,2]$.
To justify passage to the limit in the diagonal sum, choose a
continuous majorant of the strip $|Q-1|\le\varepsilon$, supported in
$|Q-1|<2\varepsilon$. Lemma~\ref{lem:gram-diagonal} applies to this majorant: after
normalization its limiting contribution tends to zero as
$\varepsilon\downarrow0$, because $Q=1$ is a hyperplane of
Lebesgue measure zero in the nine-edge coordinates. The restriction
$Q\le1+\delta$ makes the relevant region compact in combination
with the four profile support bounds. Thus, uniformly for bounded
$z,w$ and $\tau\in[1,2]$, the smoothing may be replaced in the
normalized diagonal sum by the sharp cutoff $Q\le1$.
Apply Lemma~\ref{lem:gram-diagonal} with the three indices on the $s$-side ordered as
$(m,h_1,h_2)$ and those on the $\bar s$-side as $(n,k_1,k_2)$.
Write $X=(x_{ij})_{1\le i,j\le3}$ for the nine nonnegative edge
variables, and put
\[
\begin{aligned}
a&=\sum_jx_{2j},& b&=\sum_jx_{3j},
&c&=\sum_ix_{i2},& d&=\sum_ix_{i3},\\
R_0&=\sum_jx_{1j},& C_0&=\sum_ix_{i1}.
\end{aligned}
\]
Here $R_0$ and $C_0$ are the limiting logarithmic lengths of $m$
and $n$, respectively. The diagonal cutoff becomes $R_0+C_0\le1$.
The direct term has limiting shifted weight
$\exp(z(\tfrac12-R_0)+w(\tfrac12-C_0))$; the dual term has
$\exp(z(C_0-\tfrac12)+w(R_0-\tfrac12))$.
Since the weight is now independent of $\tau$, integrating in time
gives $T\widehat\Phi(0)$. Consequently,
\[
\begin{aligned}
\frac{I(z,w)}{a_3TL^9}\longrightarrow
\widehat\Phi(0)
\int_{\substack{x_{ij}\ge0\\R_0+C_0\le1}}
&f(a)g(b)h(c)k(d)\\[-2pt]
&\ \cdot\Bigl[
e^{z(1/2-R_0)+w(1/2-C_0)}
+e^{z(C_0-1/2)+w(R_0-1/2)}\Bigr]
\prod_{i,j=1}^3\mathrm{d}x_{ij}.
\end{aligned}
\]
It remains to identify the integral on the right with the proposed
balanced kernel. Set
\[
r_{11}=x_{22},\quad r_{12}=x_{23},\quad
r_{21}=x_{32},\quad r_{22}=x_{33},\quad u=x_{11},
\]
and write $t_r=\sum_{i,j=1}^2r_{ij}$. The nine-edge matrix is then
\[
X=
\begin{pmatrix}
u&c-r_{11}-r_{21}&d-r_{12}-r_{22}\\
a-r_{11}-r_{12}&r_{11}&r_{12}\\
b-r_{21}-r_{22}&r_{21}&r_{22}
\end{pmatrix}.
\]
The nonnegativity of its entries is precisely
$r\in\mathcal R(a,b;c,d)$ together with $u\ge0$; the change from
the nine entries of $X$ to $(a,b,c,d,r_{11},r_{12},r_{21},r_{22},u)$
has Jacobian of absolute value one. Moreover,
\[
R_0=u+c+d-t_r,\qquad C_0=u+a+b-t_r.
\]
It follows that
\[
R_0+C_0\le1
\quad\Longleftrightarrow\quad
0\le u\le\frac{1-a-b-c-d+2t_r}{2}=w_r.
\]
The support condition $a+b+c+d\le S_*<1$ ensures $w_r>0$ on the
relevant region.
For the direct term, make the substitution $x=u-w_r$. Then
$-w_r\le x\le0$, and with
$\Delta=(a+b-c-d)/2$ we have
\[
\frac12-R_0=\Delta-x,\qquad
\frac12-C_0=-\Delta-x.
\]
Thus its contribution is
\[
\int_{\mathcal R(a,b;c,d)}\int_{-w_r}^0
e^{z(\Delta-x)+w(-\Delta-x)}\,\mathrm{d}x\,\mathrm{d}r.
\]
For the dual term, instead put $x=w_r-u$. Then
$0\le x\le w_r$, and
\[
C_0-\frac12=\Delta-x,\qquad
R_0-\frac12=-\Delta-x.
\]
Its contribution is therefore
\[
\int_{\mathcal R(a,b;c,d)}\int_0^{w_r}
e^{z(\Delta-x)+w(-\Delta-x)}\,\mathrm{d}x\,\mathrm{d}r.
\]
Adding the direct and dual contributions gives the uniform shifted
limit
\[
\frac{I(z,w)}{a_3TL^9}\longrightarrow
\widehat\Phi(0)
\int_{\mathbb R_{\ge0}^4}f(a)g(b)h(c)k(d)
\int_{\mathcal R(a,b;c,d)}\int_{-w_r}^{w_r}
e^{z(\Delta-x)+w(-\Delta-x)}\,\mathrm{d}x\,\mathrm{d}r\,
\mathrm{d}a\,\mathrm{d}b\,\mathrm{d}c\,\mathrm{d}d.
\]
All the estimates above are uniform on fixed complex discs in
$(z,w)$. Both sides are holomorphic in the shifts, so Cauchy's
integral formula allows us to apply the fixed differential operators
$p(\partial_z)q(\partial_w)$ and then set $z=w=0$. Since
\[
\left.p(\partial_z)q(\partial_w)
e^{z(\Delta-x)+w(-\Delta-x)}\right|_{z=w=0}
=p(\Delta-x)q(-\Delta-x),
\]
we obtain exactly the kernel $B_{p,q}(a,b;c,d)$ in \emph{\eqref{eq:gram-balanced-kernel}},
and hence the asserted limit. This completes the proof.
\end{proof}
\subsubsection{The unbalanced kernel}
For pairs of tests from different families, the terms \(H\overline J\) in \eqref{eq:projection} take the form
\[
\begin{aligned}
U_{p;f,g}\overline{V_{q;h,k}}
&=\chi(s)C_p(s)C_q(\bar s)\,
A_f(s)A_g(\bar s)A_h(\bar s)A_k(\bar s),\\[4pt]
V_{p;f,g}\overline{U_{q;h,k}}
&=\chi(s)^{-1}C_p(s)C_q(\bar s)\,
A_f(s)A_g(s)A_h(\bar s)A_k(s).
\end{aligned}
\]
Unlike the balanced case, these products contain three amplifier factors on one side and only one on the other. Using the functional equation, we reduce them to unbalanced moments, which we evaluate next.
For $a,b,c,d\geq0$, put
\[
H=a+b+c-d,\qquad
\tau=\frac{1-H}{2}.
\]
If $H\leq0$, define $M_{p,q}(a,b,c;d)=0$.
For $H>0$, let
\[
\mathcal S(a,b,c;d)=
\left\{
(r_1,r_2)\in\mathbb R_{\geq0}^2:
\begin{aligned}
r_1&\leq a,\\
r_2&\leq b,\\
0\leq r_3:=H-r_1-r_2&\leq c
\end{aligned}
\right\}.
\]
Then define
\begin{equation}\label{eq:gram-unbalanced-kernel}
\begin{split}
M_{p,q}(a,b,c;d)
={}&\int_{\mathcal S(a,b,c;d)}
\int_{\prod_{i=1}^3[-r_i/2,r_i/2]}
p(y_1+y_2+y_3-\tau)\\
&\qquad\cdot q(y_1+y_2+y_3+\tau)
\,\dd\mathbf y\,\dd r_1\,\dd r_2.
\end{split}
\end{equation}
\begin{proposition}[Unbalanced moment]\label{prop:gram-unbalanced}
The normalized integral
\[
\frac1{a_3TL^9}\int_{\R}\Phi(t/T)\chi(s)^{-1}
C_p(s)C_q(\bar s)A_f(s)A_g(s)A_h(s)A_k(\bar s)\,\dd t
\]
converges to
\[
\widehat\Phi(0)\int f(a)g(b)h(c)k(d)
M_{p,q}(a,b,c;d)\,\dd a\,\dd b\,\dd c\,\dd d.
\]
\end{proposition}
\begin{proof}
We begin by introducing two complex shifts to handle the centered
derivatives. For $z,w$ in fixed complex discs, define
\[
\begin{aligned}
I(z,w)
:= \int_{\mathbb R}
&\Phi(t/T)\chi(s)^{-1}e^{(z+w)/2}
\zeta\left(s+\frac{z}{L}\right)
\zeta\left(\bar s+\frac{w}{L}\right)\\
&\qquad\cdot
A_f(s)A_g(s)A_h(s)A_k(\bar s)\,\mathrm{d}t.
\end{aligned}
\]
By the definition of the centered derivatives, the integral in
Proposition~\ref{prop:gram-unbalanced} is precisely
\[
p(\partial_z)q(\partial_w)I(z,w)\big|_{z=w=0}.
\]
As in the balanced case, we first establish a locally uniform
asymptotic formula for $I(z,w)$ and apply the differential operators
at the end.
The main difference from the balanced case is that the factor
$\chi(s)^{-1}$ allows us to move the first zeta factor to the
same side as the second. Indeed, the centered functional equation
from Section~\ref{sec:setup} gives
\[
\chi(s)^{-1}e^{z/2}
\zeta\left(s+\frac{z}{L}\right)
=
R_T(t,z)e^{-z/2}
\zeta\left(\bar s-\frac{z}{L}\right).
\]
Consequently,
\[
\begin{aligned}
I(z,w)=\int_{\mathbb R}
&\Phi(t/T)R_T(t,z)e^{(w-z)/2}
\zeta\left(\bar s-\frac{z}{L}\right)
\zeta\left(\bar s+\frac{w}{L}\right)\\
&\qquad\cdot
A_f(s)A_g(s)A_h(s)A_k(\bar s)\,\mathrm{d}t.
\end{aligned}
\]
We retain the factor $R_T(t,z)$ throughout the moment
calculation. By Stirling's formula, uniformly for $T\le t\le 2T$
and $z$ in a fixed disc,
\[
R_T(t,z)
=
\left(\frac{t}{2\pi T}\right)^{-z/L}
+O_R(T^{-1}).
\]
The differentiated form of Stirling's formula also gives
\[
\partial_t^j R_T(t,z)\ll_{j,R}T^{-j}
\]
for every fixed integer $j\ge 0$. These estimates will allow
us to integrate by parts in the off-diagonal terms. On the
diagonal, Lemma~\ref{lem:reflection-factor} gives the simpler approximation
\[
R_T(t,z)=1+O_R(L^{-1}).
\]
Since both zeta factors now occur on the negative side, we
can replace their product by a single, sufficiently long
Dirichlet series. Put
\[
w_1=\bar s-\frac{z}{L},
\qquad
w_2=\bar s+\frac{w}{L},
\]
and define
\[
D_T(t;z,w)
=
\sum_{m,n\ge1}
\frac{e^{-mn/T^3}}{m^{w_1}n^{w_2}}.
\]
We claim that, for every fixed $R>0$ and sufficiently small
$\varepsilon>0$,
\[
\zeta(w_1)\zeta(w_2)
=
D_T(t;z,w)
+O_{\varepsilon,R}(T^{-1/2+\varepsilon}),
\]
uniformly for $T\le t\le 2T$ and $|z|,|w|\le R$.
To establish this, Mellin inversion gives
\[
D_T(t;z,w)
=
\frac{1}{2\pi i}
\int_{(2)}
\Gamma(u)T^{3u}
\zeta(w_1+u)\zeta(w_2+u)\,\mathrm{d}u.
\]
We truncate the integral at $|\Im u|\le T/4$ and shift
the contour to
\[
\Re u=-\frac12-\varepsilon,
\qquad 0<\varepsilon<\frac14.
\]
The poles of the two zeta factors occur at $u=1-w_1$
and $u=1-w_2$. Their imaginary parts are
$t+O_R(L^{-1})$, so both lie outside the rectangle
used for the contour shift. Moreover, since
$-\frac12-\varepsilon>-1$, the only pole of
$\Gamma(u)$ crossed during the shift is its simple
pole at $u=0$.
The residue at $u=0$ is
\[
\zeta(w_1)\zeta(w_2).
\]
By Stirling's formula, the horizontal integrals and
the discarded tails on the original line are
exponentially small. It remains to estimate the
integral on the new vertical line.
On this line, for $|\Im u|\le T/4$, the imaginary
parts of $w_1+u$ and $w_2+u$ have absolute value
comparable to $T$, while
\[
\Re(w_j+u)=-\varepsilon+O_R(L^{-1}).
\]
Applying the functional equation to each zeta factor
and using the absolutely convergent Dirichlet series
in the half-plane $\Re s>1$, we obtain
\[
|\zeta(w_1+u)\zeta(w_2+u)|
\ll_{\varepsilon,R}T^{1+2\varepsilon}.
\]
Since
\[
|T^{3u}|=T^{-3/2-3\varepsilon}
\]
and $\Gamma(u)$ is absolutely integrable along this
vertical line, the remaining integral is
$O_{\varepsilon,R}(T^{-1/2+\varepsilon})$.
This proves the claimed replacement.
We next verify that its error is negligible in the
moment integral. Since the four profiles are bounded
and supported in $[0,A_1],\ldots,[0,A_4]$, respectively,
\[
\begin{aligned}
|A_f(s)A_g(s)A_h(s)A_k(\bar s)|
&\ll
\prod_{j=1}^4
\sum_{n\le T^{A_j}}n^{-1/2}\\
&\ll T^{S_*/2},
\end{aligned}
\]
where $S_*=A_1+A_2+A_3+A_4<1$.
The factor $R_T(t,z)e^{(w-z)/2}$ is uniformly
bounded on fixed shift discs. Thus the integrated
replacement error is
\[
O_{\varepsilon,R}
\left(T^{1/2+S_*/2+2\varepsilon}\right).
\]
Choosing
\[
0<\varepsilon<\frac{1-S_*}{4}
\]
makes the exponent strictly smaller than one.
In particular, the error is $o(T)$, and hence
$o(TL^9)$, uniformly on fixed shift discs.
We now insert the smoothed series into $I(z,w)$
and expand the four Dirichlet polynomials. Let
$\ell_1,\ell_2,\ell_3,\ell_4$ denote their summation
indices, and write
\[
P=\ell_1\ell_2\ell_3,
\qquad
Q=\ell_4mn.
\]
For brevity, put
\[
\begin{aligned}
F_T(\ell_1,\ell_2,\ell_3,\ell_4)
={}&f\left(\frac{\log\ell_1}{L}\right)
g\left(\frac{\log\ell_2}{L}\right)\\
&\cdot h\left(\frac{\log\ell_3}{L}\right)
k\left(\frac{\log\ell_4}{L}\right).
\end{aligned}
\]
The resulting expression is
\[
\begin{aligned}
I(z,w)
={}&e^{(w-z)/2}
\sum_{\ell_1,\ldots,\ell_4,m,n\ge1}
\frac{F_T(\ell_1,\ldots,\ell_4)}
{\sqrt{PQ}}\,
m^{z/L}n^{-w/L}e^{-mn/T^3}\\
&\quad\cdot
\int_{\mathbb R}
\Phi(t/T)R_T(t,z)
\exp\left(it\log\frac{Q}{P}\right)\,\mathrm{d}t
+o(T).
\end{aligned}
\]
We separate the diagonal $P=Q$ from the
off-diagonal $P\ne Q$. The diagonal equation is
\[
\ell_1\ell_2\ell_3=\ell_4mn.
\]
We first show that the off-diagonal contribution is
negligible. Write
\[
A_+=A_1+A_2+A_3.
\]
Since $\ell_j\le T^{A_j}$, we have
$P\le T^{A_+}$. If $P\ne Q$, the integrality of
$P$ and $Q$ implies
\[
\left|\log\frac{Q}{P}\right|
\gg T^{-A_+}.
\]
Indeed, when $P$ and $Q$ are comparable, this
follows from $|P-Q|\ge1$; otherwise the logarithm
is bounded away from zero.
Now $\Phi$ is smooth and compactly supported in
$(1,2)$, and the derivative estimates for
$R_T(t,z)$ imply
\[
\frac{\mathrm{d}^j}{\mathrm{d}t^j}
\bigl(\Phi(t/T)R_T(t,z)\bigr)
\ll_{j,R,\Phi}T^{-j}.
\]
Repeated integration by parts therefore gives,
for every fixed integer $N\ge1$,
\[
\begin{aligned}
&\int_{\mathbb R}
\Phi(t/T)R_T(t,z)
\exp\left(it\log\frac{Q}{P}\right)\,\mathrm{d}t\\
&\qquad\ll_{N,R,\Phi}
T\left(T\left|\log\frac{Q}{P}\right|\right)^{-N}\\
&\qquad\ll_{N,R,\Phi}
T^{1-N(1-A_+)}
\end{aligned}
\]
whenever $P\ne Q$.
It remains to sum this bound over the indices.
The exponential smoothing ensures that the
absolute sum of the coefficients grows at most
polynomially in $T$. More precisely, for every
fixed $\delta>0$,
\[
\sum_{m,n\ge1}
\frac{e^{-mn/T^3}|m^{z/L}n^{-w/L}|}{\sqrt{mn}}
\ll_{R,\delta}T^{3/2+\delta},
\]
uniformly on fixed shift discs. Combining this
with the amplifier bound gives a total coefficient
mass
\[
O_{R,\delta}\left(
T^{3/2+S_*/2+\delta}
\right).
\]
Hence the off-diagonal contribution is
\[
O_{N,R,\delta}\left(
T^{1+3/2+S_*/2+\delta-N(1-A_+)}
\right).
\]
Since $A_+<1$, we may choose $N$ sufficiently
large to make this $o(T)$. Thus only the diagonal
contributes to the leading term.
On the diagonal $P=Q$, we have
\[
mn=\frac{P}{\ell_4}\le P\le T^{A_+}.
\]
Consequently,
\[
e^{-mn/T^3}
=
1+O\left(T^{A_+-3}\right)
=
1+o(1)
\]
uniformly over the diagonal summation indices.
Furthermore, Lemma~\ref{lem:reflection-factor} gives
\[
\begin{aligned}
\int_{\mathbb R}\Phi(t/T)R_T(t,z)\,\mathrm{d}t
&=
\int_{\mathbb R}\Phi(t/T)\,\mathrm{d}t
+O_{R,\Phi}(T/L)\\
&=
T\widehat{\Phi}(0)+O_{R,\Phi}(T/L).
\end{aligned}
\]
The absolute diagonal sum is $O(L^9)$, as follows
from Lemma~\ref{lem:gram-diagonal} and the compact support of the
profiles. Since the shift-dependent factors
$m^{z/L}$ and $n^{-w/L}$ are uniformly bounded
on the diagonal, both replacements above produce
an error $o(TL^9)$.
We therefore obtain
\[
\begin{aligned}
I(z,w)
={}&T\widehat{\Phi}(0)e^{(w-z)/2}
\sum_{\substack{\ell_1\ell_2\ell_3=\ell_4mn}}
\frac{F_T(\ell_1,\ldots,\ell_4)}
{\ell_1\ell_2\ell_3}\,
m^{z/L}n^{-w/L}\\
&\quad+o(TL^9),
\end{aligned}
\]
uniformly for $z,w$ in fixed complex discs.
We now evaluate this diagonal sum using
Lemma~\ref{lem:gram-diagonal}. Its three positive-side variables
$\ell_1,\ell_2,\ell_3$ correspond to the three
row sums of the nine-edge matrix. The three
negative-side variables $m,n,\ell_4$ correspond
to its column sums.
Fix row sums $a,b,c$ and let $d$ be the
column sum corresponding to $\ell_4$. Put
\[
H=a+b+c-d.
\]
We parameterize the nine nonnegative edge
variables by the matrix
\[
\begin{pmatrix}
v_1&r_1-v_1&a-r_1\\
v_2&r_2-v_2&b-r_2\\
v_3&r_3-v_3&c-r_3
\end{pmatrix},
\]
where
\[
r_1+r_2+r_3=H,
\qquad
0\le r_1\le a,\quad
0\le r_2\le b,\quad
0\le r_3\le c,
\]
and
\[
0\le v_i\le r_i
\qquad (i=1,2,3).
\]
The first two column sums are
\[
v_1+v_2+v_3
\quad\text{and}\quad
H-v_1-v_2-v_3,
\]
while the third column sum is $d$, as required.
Eliminating
\[
r_3=H-r_1-r_2
\]
leaves precisely the region
$\mathcal S(a,b,c;d)$ defined above. The change of
variables from the nine edge coordinates to
\[
(a,b,c,d,r_1,r_2,v_1,v_2,v_3)
\]
has Jacobian of absolute value one.
The shift-dependent factors in the diagonal
sum consequently become
\[
m^{z/L}n^{-w/L}
\longrightarrow
\exp\left(
z(v_1+v_2+v_3)
-w(H-v_1-v_2-v_3)
\right).
\]
Including the centering factor $e^{(w-z)/2}$,
the limiting shifted weight is
\[
\exp\left(
z\left(v_1+v_2+v_3-\frac12\right)
+
w\left(v_1+v_2+v_3-H+\frac12\right)
\right).
\]
For $H>0$, Lemma~\ref{lem:gram-diagonal} therefore yields
\[
\begin{aligned}
\lim_{T\to\infty}
\frac{I(z,w)}{a_3TL^9}
={}&\widehat{\Phi}(0)
\int_{\mathbb R_{\ge0}^4}
f(a)g(b)h(c)k(d)\\
&\quad\cdot
\int_{\mathcal S(a,b,c;d)}
\int_{\prod_{i=1}^3[0,r_i]}
\exp\left(
z\left(\sum_{i=1}^3v_i-\frac12\right)
\right)\\
&\qquad\qquad\cdot
\exp\left(
w\left(\sum_{i=1}^3v_i-H+\frac12\right)
\right)
\,\mathrm{d}v\,\mathrm{d}r_1\,\mathrm{d}r_2\,
\mathrm{d}a\,\mathrm{d}b\,\mathrm{d}c\,\mathrm{d}d.
\end{aligned}
\]
When $H<0$, the corresponding region is empty;
when $H=0$, it has measure zero. Thus both cases
give zero, in agreement with the definition
of $M_{p,q}$.
We finally apply the differential operators
$p(\partial_z)$ and $q(\partial_w)$.
All estimates above are uniform on fixed
complex shift discs. Moreover, on the diagonal
the logarithmic variables are confined to a
fixed compact region, so the shifted diagonal
weights are holomorphic and uniformly bounded
on such discs. The convergence furnished by
Lemma~\ref{lem:gram-diagonal} is therefore locally uniform in $z,w$.
Cauchy's integral formula justifies applying
the differential operators to the limiting
integral.
Writing
\[
V=v_1+v_2+v_3,
\]
we obtain the polynomial weight
\[
p\left(V-\frac12\right)
q\left(V-H+\frac12\right).
\]
It follows that the limiting kernel before
the final change of variables is
\[
\begin{aligned}
\int_{\mathcal S(a,b,c;d)}
\int_{\prod_{i=1}^3[0,r_i]}
&p\left(\sum_{i=1}^3v_i-\frac12\right)\\
&\cdot q\left(\sum_{i=1}^3v_i-H+\frac12\right)
\,\mathrm{d}v\,\mathrm{d}r_1\,\mathrm{d}r_2.
\end{aligned}
\]
To obtain the symmetric form stated in
the proposition, put
\[
y_i=v_i-\frac{r_i}{2},
\qquad i=1,2,3.
\]
Then
\[
-\frac{r_i}{2}\le y_i\le\frac{r_i}{2},
\qquad
\sum_{i=1}^3v_i
=
\sum_{i=1}^3y_i+\frac H2.
\]
Recalling that
\[
\tau=\frac{1-H}{2},
\]
the two polynomial arguments become
\[
\sum_{i=1}^3v_i-\frac12
=
\sum_{i=1}^3y_i-\tau
\]
and
\[
\sum_{i=1}^3v_i-H+\frac12
=
\sum_{i=1}^3y_i+\tau.
\]
The change of variables has Jacobian one.
Thus the limiting kernel is exactly
\[
\begin{aligned}
M_{p,q}(a,b,c;d)
={}&\int_{\mathcal S(a,b,c;d)}
\int_{\prod_{i=1}^3[-r_i/2,r_i/2]}\\
&\quad
p\left(y_1+y_2+y_3-\tau\right)
q\left(y_1+y_2+y_3+\tau\right)
\,\mathrm{d}\mathbf{y}\,\mathrm{d}r_1\,\mathrm{d}r_2,
\end{aligned}
\]
which agrees with \eqref{eq:gram-unbalanced-kernel}.
We conclude that
\[
\begin{aligned}
\lim_{T\to\infty}
\frac{1}{a_3TL^9}
\int_{\mathbb R}
&\Phi(t/T)\chi(s)^{-1}
C_p(s)C_q(\bar s)\\
&\quad\cdot
A_f(s)A_g(s)A_h(s)A_k(\bar s)\,\mathrm{d}t\\
={}&\widehat{\Phi}(0)
\int_{\mathbb R_{\ge0}^4}
f(a)g(b)h(c)k(d)
M_{p,q}(a,b,c;d)
\,\mathrm{d}a\,\mathrm{d}b\,\mathrm{d}c\,\mathrm{d}d,
\end{aligned}
\]
as claimed.
\end{proof}
\subsubsection{Reflection errors and all projected Gram blocks}
We first control the error arising from the centered
functional equation. By the functional equation,
\[
\chi(s)^{-1}C_q(s)-C_{q^-}(\bar s)
=
\left.q(\partial_z)\left[
(R_T(t,z)-1)e^{-z/2}
\zeta(\bar s-z/L)
\right]\right|_{z=0}.
\]
By Lemma~\ref{lem:reflection-factor} and Leibniz's rule, for every fixed
polynomial $q$ of degree at most four,
\begin{equation}\label{eq:gram-reflection-error}
\left|\chi(s)^{-1}C_q(s)-C_{q^-}(\bar s)\right|
\ll_q L^{-1}\sum_{j=0}^4|C_{x^j}(\bar s)|,
\end{equation}
uniformly for $T\le t\le 2T$.
We now show that the reflection error is negligible in the Gram
matrix. Multiplying \eqref{eq:gram-reflection-error} by the two
Dirichlet-polynomial factors of a test, say
$A_f(s)A_g(\bar s)$, and using
\[
\left(\sum_{j=0}^4 a_j\right)^2
\leq 5\sum_{j=0}^4 a_j^2,
\]
we obtain
\[
\begin{aligned}
&\left|
\bigl(\chi(s)^{-1}C_q(s)-C_{q^-}(\bar s)\bigr)
A_f(s)A_g(\bar s)
\right|^2\\
&\qquad\ll_q L^{-2}
\sum_{j=0}^4
\left|C_{x^j}(\bar s)A_f(s)A_g(\bar s)\right|^2.
\end{aligned}
\]
Choose a smooth nonnegative weight $\Psi$ satisfying
$\Psi\geq|\Phi|$. By Proposition~\ref{prop:gram-balanced},
each term in the sum has weighted integral $O(TL^9)$.
Consequently,
\[
\begin{aligned}
&\int_{\mathbb R}\Psi(t/T)
\left|
\bigl(\chi(s)^{-1}C_q(s)-C_{q^-}(\bar s)\bigr)
A_f(s)A_g(\bar s)
\right|^2\,\mathrm{d}t\\
&\qquad\ll L^{-2}\cdot TL^9
=O(TL^7).
\end{aligned}
\]
The other factor in the corresponding Gram product
has squared weighted norm $O(TL^9)$, again by
Proposition~\ref{prop:gram-balanced}.
Thus Cauchy--Schwarz gives
\[
\begin{aligned}
&\left|\int_{\mathbb R}\Phi(t/T)
\bigl(\chi(s)^{-1}C_q(s)-C_{q^-}(\bar s)\bigr)
C_p(s)A_f(s)A_g(\bar s)
A_h(s)A_k(\bar s)\,\mathrm{d}t\right|\\
&\qquad\ll (TL^7)^{1/2}(TL^9)^{1/2}\\
&\qquad=O(TL^8)=o(TL^9).
\end{aligned}
\]
The same argument applies to the other placements
of the amplifier factors. Hence all reflection errors
are negligible at the scale $TL^9$.
Since the estimates use a nonnegative majorant of
$|\Phi|$, no positivity assumption on $\Phi$ is needed.
Finally, we consider the reflected $VV$ product. By the
projection identity, this term initially takes the form
\[
\chi(s)^{-3}C_p(s)C_q(s)
A_f(s)A_g(s)A_h(s)A_k(s).
\]
Applying the centered functional equation to $C_q(s)$,
and using the reflection-error estimate established above,
it suffices to show that
\[
\int_{\mathbb R}\Phi(t/T)\chi(s)^{-2}
C_p(s)C_{q^-}(\bar s)
A_f(s)A_g(s)A_h(s)A_k(s)\,\mathrm{d}t
=o(TL^9).
\]
We introduce bounded complex shifts for the two zeta
factors and apply the one-sided expansion used in the
proof of Proposition~\ref{prop:gram-unbalanced}.
After expanding the four Dirichlet polynomials, the
resulting main terms contain the oscillatory factor
\[
\chi(s)^{-1}
\exp\left(
it\log\frac{mn}{h_1h_2h_3h_4}
\right),
\]
where $m,n\geq 1$ and $h_i\leq T^{A_i}$.
The derivative of its phase is
\[
\begin{aligned}
\varphi'(t)
&=\log\frac{t}{2\pi}
+\log\frac{mn}{h_1h_2h_3h_4}
+O(t^{-1})\\
&\geq (1-S_*)L+O(1),
\end{aligned}
\]
since $h_1h_2h_3h_4\leq T^{S_*}$ and $mn\geq 1$.
As $S_*<1$, we therefore have
\[
\varphi'(t)\gg L
\]
uniformly for $T\leq t\leq 2T$. In particular,
the phase has no stationary point.
Writing $t=Tx$ and absorbing the shift-dependent
powers of $x$ into the smooth amplitude, repeated
integration by parts gives, for every fixed integer
$N\geq 1$,
\[
\int_{\mathbb R}\Phi(t/T)\chi(s)^{-1}
\exp\left(
it\log\frac{mn}{h_1h_2h_3h_4}
\right)
(t/T)^\tau\,\mathrm{d}t
\ll_{N,R} T^{1-N},
\]
uniformly in the summation indices and for
$|\tau|\leq R$, with $R$ fixed. Here the uniform
bounds for the derivatives of the reciprocal phase
derivative ensure that the integrations by parts
can be repeated.
The absolute sum of the coefficients in the
one-sided expansion is bounded by
$O(T^{6+S_*/2})$. Taking $N=9$, the total
contribution of the oscillatory main terms is
therefore
\[
O(T^{6+S_*/2}T^{-8})
=O(T^{-2+S_*/2}).
\]
It remains to control the error in the one-sided
replacement. The uniform replacement estimate
used in the proof of
Proposition~\ref{prop:gram-unbalanced}
bounds its integrated contribution by
\[
O\bigl(T^{1/2+2\varepsilon+S_*/2}\bigr).
\]
Choosing $\varepsilon=(1-S_*)/8$ makes the exponent
equal to $(3+S_*)/4<1$. Hence this contribution
is also $o(TL^9)$.
All estimates are uniform on fixed shift discs,
so the differential operators defining $C_p$
and $C_{q^-}$ may be applied afterwards.
Consequently, the reflected $VV$ contribution
is $o(TL^9)$ and does not enter the limiting
Gram matrix.
We can now assemble the Gram kernels for all pairs of projected
tests. For two tests with polynomial and profile variables
$(p;a,b)$ and $(q;c,d)$, respectively, define
\begin{equation}\label{eq:gram-blocks}
\begin{aligned}
G_{UU}(p,q;a,b,c,d)
&=\frac12\Bigl(
B_{p,q}(a,d;b,c)
+B_{p,q^-}(a,c;b,d)
\Bigr),\\
G_{UV}(p,q;a,b,c,d)
&=\frac12\Bigl(
M_{q,p}(b,c,d;a)
+M_{p,q^-}(a,c,d;b)
\Bigr),\\
G_{VV}(p,q;a,b,c,d)
&=\frac12 B_{p,q}(a,b;c,d),\\
G_{VU}(p,q;a,b,c,d)
&=G_{UV}(q,p;c,d,a,b).
\end{aligned}
\end{equation}
To explain these formulas, consider first the $UU$ block.
The ordinary product is
\[
\begin{aligned}
U_{p;f,g}\overline{U_{q;h,k}}
={}&C_p(s)C_q(\bar s)\\
&\cdot A_f(s)A_k(s)
A_g(\bar s)A_h(\bar s).
\end{aligned}
\]
By Proposition~\ref{prop:gram-balanced}, this contributes
$B_{p,q}(a,d;b,c)$. For the reflected product, the
centered functional equation replaces
$\chi(s)^{-1}C_q(s)$ by $C_{q^-}(\bar s)$, up to
the negligible error established above. The resulting
balanced moment contributes $B_{p,q^-}(a,c;b,d)$.
For the $UV$ block, conjugating the ordinary product
gives an unbalanced moment with profile ordering
$(b,c,d;a)$ and polynomials $(q,p)$. The reflected
product, after applying the centered functional equation,
gives the ordering $(a,c,d;b)$ with polynomials
$(p,q^-)$. These yield the two terms in $G_{UV}$.
For the $VV$ block, the ordinary product is balanced,
while the reflected contribution is $o(TL^9)$ by the
preceding nonstationary phase argument. The $VU$
formula follows by symmetry. In each case, the factor
$1/2$ comes from the projection identity~\eqref{eq:projection}.
Finally, for admissible profiles $f,g,h,k$, the
corresponding Gram coefficient is obtained by integrating
the appropriate point kernel:
\[
\int_{\mathbb R_{\geq0}^4}
f(a)g(b)h(c)k(d)\,
G_{\tau\sigma}(p,q;a,b,c,d)
\,\mathrm{d}a\,\mathrm{d}b\,\mathrm{d}c\,\mathrm{d}d,
\]
where $\tau,\sigma\in\{U,V\}$ denote the two test
families. This determines all entries of the
denominator Gram matrix.
\section{Assembly of the Lower Bound}
\subsection{Approximation by admissible profiles}
\label{sec:approximation}
The moment formulas established in Section~\ref{sec:analytic} express the
numerator and denominator of our test combinations as
integrals of explicit kernels against admissible profiles.
For the construction of a rational certificate, however,
it is more convenient to evaluate these kernels directly
at fixed points. We now show that such point evaluations
can be approximated by genuine admissible profiles, so
that the resulting finite-dimensional bounds remain
applicable to the sixth moment.
Consider a finite collection of endpoint tests, indexed
by $j=1,\ldots,m$. Each test is specified by a type
$U$ or $V$, a polynomial $p_j$, and two endpoints
$a_j,b_j$. For a $U$-test, we require
\[
00$, the associated endpoint quotient is
\[
R=\frac{9!N^2}{D}.
\]
Our purpose is to show that this quotient is the limit
of the corresponding quotients obtained from
admissible test functions.
\begin{proposition}[Approximation of endpoint tests]
\label{prop:profile-approximation}
Let a finite collection of endpoint tests satisfy the
conditions above, and let $c\in\mathbb R^m$ be fixed.
There exists a sequence of collections of admissible
$U$- and $V$-tests such that their numerator vectors
$k^{(n)}$ and projected Gram matrices $G^{(n)}$
satisfy
\[
k_j^{(n)}\longrightarrow k_j,
\qquad
G_{ij}^{(n)}\longrightarrow G_{ij}
\]
for every $i,j$.
Consequently, with
\[
N_n=c^\top k^{(n)},\qquad
D_n=c^\top G^{(n)}c,
\]
we have
\[
N_n\longrightarrow N,\qquad
D_n\longrightarrow D.
\]
If $D>0$, then $D_n>0$ for all sufficiently large
$n$, and
\begin{equation}\label{eq:profile-limits}
\frac{9!N_n^2}{D_n}
\longrightarrow
\frac{9!N^2}{D}.
\end{equation}
\end{proposition}
\begin{proof}
We construct the approximating profiles separately
for the two test families. The constructions differ
because the $U$-family requires smooth compactly
supported profiles, whereas the $V$-family requires
polynomial profiles on a common interval.
Fix a nonnegative function
$\psi\in C_c^\infty((0,1))$ satisfying
\[
\int_{\mathbb R}\psi(x)\,\mathrm{d}x=1.
\]
Let $h_n>0$ decrease to zero, and for an endpoint
$a>0$ define
\[
f_{a,n}(x)
=\frac{1}{h_n}
\psi\left(\frac{a-x}{h_n}\right).
\]
We choose $h_n0$, its integration region in the first
two variables is contained in
\[
\bigl\{(r_1,r_2)\in\mathbb R_{\geq0}^2:
r_1+r_2\leq H\bigr\}.
\]
As $H$ decreases to zero, this region shrinks
to a set of measure zero. Since the integrand
is locally bounded, the integral tends to zero.
Thus $M_{p,q}$ is continuous also across $H=0$.
Finally, all projected Gram kernels are finite
linear combinations of these balanced and
unbalanced kernels, with the polynomial
reflections and permutations established in
Section~\ref{sec:denominator}. They are therefore continuous
in their endpoint variables as well.
\medskip
For each fixed $n$, let $k^{(n)}$ and $G^{(n)}$
be the normalized limiting numerator vector
and projected Gram matrix obtained by applying
the moment formulas of Section~\ref{sec:analytic} to the
admissible profiles constructed above.
For example, the numerator coefficient of a
$U$-test is
\[
k_j^{(n)}
=\int_{\mathbb R^2}
f_{a_j,n}(x)f_{b_j,n}(y)
K_{p_j}(x,y)\,\mathrm{d}x\,\mathrm{d}y,
\]
whereas for a $V$-test it is
\[
k_j^{(n)}
=\int_{\mathbb R^2}
\beta_{a_j,n}(x)\beta_{b_j,n}(y)
K_{p_j}(x,1-y)\,\mathrm{d}x\,\mathrm{d}y.
\]
The Gram entries $G_{ij}^{(n)}$ are obtained
analogously by integrating the corresponding
projected point kernels against the four
profile measures belonging to tests $i$ and $j$.
All these measures are probability measures
supported in fixed compact intervals.
Their weak convergence to the respective
point masses implies weak convergence of
their finite product measures. By the
continuity of the kernels established above,
we consequently obtain
\[
k_j^{(n)}\longrightarrow k_j,
\qquad
G_{ij}^{(n)}\longrightarrow G_{ij}
\]
for every pair of indices $i,j$.
Since the collection of tests is finite,
these convergences hold simultaneously
for all entries. For the fixed coefficient
vector $c$, it follows that
\[
\begin{aligned}
N_n
&=\sum_{j=1}^m c_jk_j^{(n)}
\longrightarrow
\sum_{j=1}^m c_jk_j=N,\\
D_n
&=\sum_{i,j=1}^m c_ic_jG_{ij}^{(n)}
\longrightarrow
\sum_{i,j=1}^m c_ic_jG_{ij}=D.
\end{aligned}
\]
If $D>0$, the convergence $D_n\to D$
implies $D_n>0$ for all sufficiently large
$n$. Hence
\[
\frac{9!N_n^2}{D_n}
\longrightarrow
\frac{9!N^2}{D},
\]
as claimed.
Notice that we only require positivity of
the quadratic form in the chosen direction
$c$. No invertibility or positive definiteness
of the entire Gram matrix is needed.
\end{proof}
\subsection{From endpoint kernels to the sixth moment}
\label{sec:passage}
We now use the moment formulas from Section~\ref{sec:analytic}
and the Cauchy--Schwarz inequality from Section~\ref{sec:combinations} to
convert this endpoint quotient into a lower bound for the
sixth moment.
The moment formulas initially provide estimates with a
smooth weight supported on an interval of the form
$(Y,2Y)$. To obtain a bound for the full moment $M_3(T)$,
we shall combine finitely many such estimates over
disjoint dyadic intervals.
\begin{proposition}[Endpoint-to-moment transfer]
\label{prop:endpoint-transfer}
Let a finite collection of endpoint tests and a real
coefficient vector $c$ satisfy the conditions of
Proposition~\ref{prop:profile-approximation}. Let
$k$ and $G$ be the corresponding endpoint numerator
vector and projected Gram matrix, and put
\[
N=c^\top k,\qquad D=c^\top Gc.
\]
If $D>0$, then
\begin{equation}\label{eq:endpoint-moment-bound}
\liminf_{T\to\infty}
\frac{M_3(T)}
{(a_3/9!)T(\log T)^9}
\geq \frac{9!N^2}{D}.
\end{equation}
\end{proposition}
\begin{proof}
Write
\[
R=\frac{9!N^2}{D}.
\]
If $N=0$, then $R=0$ and the assertion follows
immediately from the nonnegativity of $M_3(T)$.
We may therefore assume $R>0$.
By Proposition~\ref{prop:profile-approximation},
there are admissible profile collections with
associated numerator and denominator coefficients
$N_n,D_n$ satisfying
\[
N_n\longrightarrow N,\qquad
D_n\longrightarrow D>0.
\]
In particular, $D_n>0$ for all sufficiently large
$n$, and the corresponding quotients
\[
R_n=\frac{9!N_n^2}{D_n}
\]
satisfy
\begin{equation}\label{eq:transfer-profile-ratio}
R_n\longrightarrow R.
\end{equation}
We first establish a weighted lower bound for each
sufficiently large fixed $n$.
Let $\Phi\in C_c^\infty((1,2))$ be nonnegative, with
$\widehat{\Phi}(0)=\int_{\mathbb R}\Phi(x)\,\mathrm{d}x>0$.
At scale $Y$, let $F_{c,n,Y}$ denote the projected
linear combination from Section~\ref{sec:combinations}, formed using
the admissible profiles of index $n$.
For the remainder of this step, $n$, the test
polynomials, their coefficients, and $\Phi$ are
all fixed while $Y\to\infty$.
The numerator moment formulas of Section~\ref{sec:analytic} give
\begin{equation}\label{eq:transfer-numerator}
\frac{1}{a_3Y(\log Y)^9}
\int_{\mathbb R}
\Phi(t/Y)\zeta(\tfrac12+it)^3
\overline{F_{c,n,Y}(t)}\,\mathrm{d}t
\longrightarrow
\widehat{\Phi}(0)N_n.
\end{equation}
Likewise, the denominator moment formulas,
including all pairwise interactions between
the projected tests, give
\begin{equation}\label{eq:transfer-denominator}
\frac{1}{a_3Y(\log Y)^9}
\int_{\mathbb R}
\Phi(t/Y)|F_{c,n,Y}(t)|^2\,\mathrm{d}t
\longrightarrow
\widehat{\Phi}(0)D_n.
\end{equation}
Since $\widehat{\Phi}(0)D_n>0$, the denominator
integral is positive for all sufficiently large
$Y$. We may therefore apply the weighted
Cauchy--Schwarz inequality:
\[
\begin{aligned}
&\int_{\mathbb R}\Phi(t/Y)
|\zeta(\tfrac12+it)|^6\,\mathrm{d}t\\
&\qquad\geq
\frac{
\left|
\int_{\mathbb R}\Phi(t/Y)
\zeta(\tfrac12+it)^3
\overline{F_{c,n,Y}(t)}\,\mathrm{d}t
\right|^2
}{
\int_{\mathbb R}\Phi(t/Y)
|F_{c,n,Y}(t)|^2\,\mathrm{d}t
}.
\end{aligned}
\]
Inserting the limits
\eqref{eq:transfer-numerator} and
\eqref{eq:transfer-denominator}, we obtain
\[
\begin{aligned}
&\liminf_{Y\to\infty}
\frac{\int_{\mathbb R}\Phi(t/Y)
|\zeta(\tfrac12+it)|^6\,\mathrm{d}t}
{(a_3/9!)Y(\log Y)^9}\\
&\qquad\geq
9!\frac{\widehat{\Phi}(0)^2N_n^2}
{\widehat{\Phi}(0)D_n}\\
&\qquad=\widehat{\Phi}(0)R_n.
\end{aligned}
\]
Thus every sufficiently large fixed profile index
$n$ gives a weighted sixth-moment lower bound
with constant $\widehat{\Phi}(0)R_n$.
We next remove the smooth weight by combining
estimates at finitely many dyadic scales.
Fix an integer $J\geq1$, and for $0\leq j0,\qquad R_n>\rho.
\]
Next, choose a smooth weight satisfying
\[
0\leq\Phi\leq1,\qquad
\Phi\in C_c^\infty((1,2)),
\]
whose integral $\widehat{\Phi}(0)$ is
sufficiently close to one. Such weights
exist, for example, by choosing $\Phi$
equal to one on
$[1+\varepsilon,2-\varepsilon]$
and letting $\varepsilon$ become small.
Finally, choose the fixed integer $J$
sufficiently large that
\[
\widehat{\Phi}(0)R_n(1-2^{-J})>\rho.
\]
The preceding lower bound then gives
\[
\liminf_{T\to\infty}
\frac{M_3(T)}
{(a_3/9!)T(\log T)^9}>\rho.
\]
Since $\rho0,\qquad \frac{9!N^2}{D}>41.008.
\]
On a subspace where the Gram matrix is positive definite,
the quotient is maximized in the direction $G^{-1}k$.
This motivates our choice of coefficients, but the proof
uses only the explicit rational certificate below.
Neither the number of tests nor the degree-four cutoff
is claimed to be optimal.
\subsubsection{Certificate data}
The certificate consists of 250 rows
\[
(\mathrm{type},A,B,r,c)
\]
in the file \texttt{certificate\_250.tsv}~\cite{certificate}. Each row specifies a family $U$ or $V$,
rational endpoints
\[
a=\frac{A}{10^8},\qquad b=\frac{B}{10^8},
\]
a rational coefficient $c$, and a test polynomial
\[
p_r(x)=T_r(2x),\qquad 0\le r\le 4,
\]
where $T_r$ is the Chebyshev polynomial of the first kind.
There are 220 $U$-tests and 30 $V$-tests.
For every $U$-test, the endpoints satisfy
$00.
\]
For the $V$-tests, all endpoints are positive and
\[
\max_{V\text{-tests}}\{a,b\}
=0.249975
<\eta=\frac{19999}{80000}<\frac14.
\]
Thus every test satisfies the admissibility conditions
of Section~\ref{sec:approximation}.
\subsubsection{Exact rational evaluation}
We now explain why the endpoint kernels can be
evaluated using exact rational arithmetic.
The integral representations from Section~\ref{sec:analytic} involve
polynomial integrands over regions defined by linear
inequalities with rational coefficients. At the
rational endpoints of our certificate, their evaluation
reduces to finitely many exact polynomial integrations.
For the numerator kernel $K_p$, the simplex
representation from Section~\ref{sec:numerator} is evaluated using
standard simplex moments. Since $\deg p\le 4$,
only moments of order at most four are needed,
followed by polynomial integration over a rational
interval.
For the balanced kernel $B_{p,q}$, we slice the
integration region according to the total edge mass
\[
t=\sum_{i,j=1}^{2}r_{ij}.
\]
The resulting slice density is piecewise polynomial.
Splitting the interval at its finitely many rational
breakpoints therefore reduces the evaluation to
ordinary polynomial integrals.
For the unbalanced kernel $M_{p,q}$, we first
integrate over the inner variables $y_i$.
The remaining polynomial integral over a capped
simplex is evaluated by inclusion--exclusion
and standard simplex moment formulas.
Together with the projected Gram formulas from
Section~\ref{sec:denominator}, these procedures determine every
entry of $k$ and $G$ exactly. The exact evaluation routines are available in the repository~\cite{certificate}.
\subsubsection{Certified inequalities and conclusion}
Let $k$ and $G$ be the endpoint values obtained
from the certificate, and let $c$ be its rational
coefficient vector. Exact rational evaluation gives
\[
N=c^\top k>46159,
\qquad
0\frac{9!\cdot 46159^2}{18854000000000}\\
&=\frac{1208080410327}{29459375000}\\
&>41.008.
\end{aligned}
\]
This proves Theorem~\ref{thm:main}.
\begin{thebibliography}{9}
\bibitem{BBLR}
S.~Bettin, H.~M.~Bui, X.~Li and M.~Radziwi\l\l,
\emph{A quadratic divisor problem and moments of the Riemann zeta-function},
J. Eur. Math. Soc. \textbf{22} (2020), 3953--3980.
\href{https://doi.org/10.4171/JEMS/999}{doi:10.4171/JEMS/999}.
The equation, proposition and page numbers used here refer to
\href{https://arxiv.org/abs/1609.02539v1}{arXiv:1609.02539v1}.
\bibitem{CG}
J.~B.~Conrey and A.~Ghosh,
\emph{A conjecture for the sixth power moment of the Riemann zeta-function},
Int. Math. Res. Not. \textbf{1998}, no.~15, 775--780.
\href{https://doi.org/10.1155/S1073792898000476}{doi:10.1155/S1073792898000476}.
\href{https://arxiv.org/abs/math/9807187}{arXiv:math/9807187}.
\bibitem{deBoor}
C.~de~Boor,
\emph{Math/CS 887: Approximation Theory}, class notes, Spring 2003,
version of 21 November 2009, pp.~52--53, formula~(83).
\url{https://pages.cs.wisc.edu/~deboor/887/notes.pdf}.
\bibitem{DP}
B.~Durkan and T.~Page,
\emph{Amplified moments of the Riemann zeta function},
\href{https://arxiv.org/abs/2606.27323v2}{arXiv:2606.27323v2} (2026).
\bibitem{certificate}
N.~Fraccaro,
\emph{Sixth moment: Lean formalization and exact certificate},
repository snapshot \texttt{950bfaff8724d4adb1c37535634a139b70228df0}.
\url{https://github.com/Niklas-Fraccaro/sixth-moment-lean/tree/950bfaff8724d4adb1c37535634a139b70228df0}.
The endpoint table is \texttt{certificate\_250.tsv}; the analytic limits
and exact evaluations are in \texttt{SixthMoment/Numerator},
\texttt{SixthMoment/Denominator} and \texttt{SixthMoment/Certificate}.
\bibitem{KS}
J.~P.~Keating and N.~C.~Snaith,
\emph{Random matrix theory and $\zeta(1/2+it)$},
Comm. Math. Phys. \textbf{214} (2000), 57--89.
\href{https://doi.org/10.1007/s002200000261}{doi:10.1007/s002200000261}.
\end{thebibliography}
\end{document}