64c3814
==== TABLE OF CONTENTS ====
The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//figures/masks.tex
The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//figures/triangle.tex
The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//main.tex
The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//preamble.tex
The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/consequences.tex
The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/heat.tex
The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/introduction.tex
The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/jumps.tex
The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/matrix.tex
The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/maximal.tex
The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/trace.tex
The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//references.bib
===========================
==== BEGIN The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//figures/masks.tex ====
\begin{figure}[htbp]
\centering
\begin{tikzpicture}[x=1cm,y=1cm,font=\small,
active/.style={line width=2pt,draw=blue!65!black},
event/.style={densely dotted,draw=black!24},
endpoint/.style={circle,draw=blue!65!black,fill=white,
line width=.65pt,inner sep=0pt,minimum size=4pt},
endpoint label/.style={font=\footnotesize,fill=white,inner sep=2pt},
energy/.style={line width=1pt,draw=black},
jump/.style={-{Stealth[length=3pt]},densely dashed,
line width=.8pt,draw=red!65!black}]
\foreach \t in {2.3,3.8,6.4,9.0,11.6}
\draw[event] (\t,.1)--(\t,6.15);
\node[anchor=west] at (1.9,6.65) {Output-dependent scale intervals};
\foreach \yy in {5.8,4.8,3.8}
\draw[black!22] (1.95,\yy)--(12.25,\yy);
\node[anchor=east] at (1.7,5.8) {$(i_1,j_1)$};
\node[anchor=east] at (1.7,4.8) {$(i_2,j_2)$};
\node[anchor=east] at (1.7,3.8) {$(i_3,j_3)$};
\draw[active] (2.3,5.8)--(6.4,5.8);
\draw[active] (3.8,4.8)--(9.0,4.8);
\draw[active] (6.4,3.8)--(11.6,3.8);
\foreach \xx/\yy in {2.3/5.8,6.4/5.8,3.8/4.8,9.0/4.8,6.4/3.8,11.6/3.8}
\node[endpoint] at (\xx,\yy) {};
\node[endpoint label,above] at (2.3,5.91) {$\varepsilon_{i_1j_1}^{\,2}$};
\node[endpoint label,above] at (6.4,5.91) {$\rho_{i_1j_1}^{\,2}$};
\node[endpoint label,above] at (3.8,4.91) {$\varepsilon_{i_2j_2}^{\,2}$};
\node[endpoint label,above] at (9.0,4.91) {$\rho_{i_2j_2}^{\,2}$};
\node[endpoint label,above] at (6.4,3.91) {$\varepsilon_{i_3j_3}^{\,2}$};
\node[endpoint label,above] at (11.6,3.91) {$\rho_{i_3j_3}^{\,2}$};
\draw[-{Stealth[length=4pt]},black!65] (1.95,3.35)--(12.6,3.35)
node[right] {$s$};
\node[anchor=west,fill=white,inner sep=2pt] at (1.9,2.9) {Integrated energy};
\draw[-{Stealth[length=4pt]},black!65] (1.95,.1)--(12.6,.1)
node[right] {$s$};
\node[anchor=east] at (1.7,1.7) {$J(s)$};
\draw[energy] (2.3,2.2) .. controls (2.9,2.05) and (3.3,1.91) .. (3.8,1.82);
\draw[jump] (3.8,1.82)--(3.8,2.38);
\draw[energy] (3.8,2.38) .. controls (4.7,2.13) and (5.4,1.85) .. (6.4,1.64);
\draw[jump] (6.4,1.64)--(6.4,1.97);
\draw[energy] (6.4,1.97) .. controls (7.2,1.75) and (8.1,1.37) .. (9.0,1.18);
\draw[jump] (9.0,1.18)--(9.0,.78);
\draw[energy] (9.0,.78) .. controls (9.9,.61) and (10.7,.43) .. (11.6,.32);
\node[below] at (2.3,.02) {$s_-$};
\node[below] at (11.6,.02) {$s_+$};
\end{tikzpicture}
\caption{Three selected output pairs and a schematic integrated energy,
not an exact heat-flow solution. The entry $A_0^s(i,j)$ is retained only
on $(\varepsilon_{ij}^2,\rho_{ij}^2)$. Between successive event scales,
$J(s)$ is nonincreasing by the fixed-array dissipation inequality.
Dashed arrows connect its one-sided limits at events; the values at the
events themselves are not shown. Jumps may have either sign, and
endpoints may coincide, as illustrated
by $\rho_{i_1j_1}^2=\varepsilon_{i_3j_3}^2$. The bound
$\sum_\tau |\Delta J(\tau)|\le C\mathcal S(A)$, together with the
initial-energy bound, controls the accumulated dissipation independently
of the number of event scales.}
\label{fig:mask-energy}
\end{figure}
==== END The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//figures/masks.tex ====
==== BEGIN The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//figures/triangle.tex ====
\begin{figure}[htbp]
\centering
\begin{tikzpicture}[>=Stealth, every node/.style={font=\small},
space/.style={circle,draw=black!60,minimum size=9mm,inner sep=2pt},
edge/.style={->,thick,blue!65!black}]
\node[space] (h0) at (0,1.9) {$H_0$};
\node[space] (h1) at (-1.8,-.8) {$H_1$};
\node[space] (h2) at (1.8,-.8) {$H_2$};
\draw[edge] (h1) -- node[left=5pt] {$W_0$} (h0);
\draw[edge] (h2) -- node[below=5pt] {$W_1$} (h1);
\draw[edge] (h0) -- node[right=5pt] {$W_2$} (h2);
\node[below=18pt] at (0,-.8) {the cyclic product $W_0W_1W_2$};
\node[space] (r1) at (5,1.3) {$H_1$};
\node[space] (r2) at (5,-.8) {$H_2$};
\node[space] (r0) at (7.5,.25) {$H_0$};
\draw[edge] (r1) -- node[above=5pt] {$W_0$} (r0);
\draw[edge] (r2) -- node[below=5pt] {$W_2^*$} (r0);
\node[below=18pt] at (6.25,-.8) {$R_0=[\,W_0\quad W_2^*\,]$};
\end{tikzpicture}
\caption{The three edge maps and the row star at vertex $0$.
Taking the adjoint of $W_2$ makes both blocks of $R_0$
have target $H_0$. The two Grams $R_0^*R_0$ and $R_0R_0^*$
have the same positive eigenvalues on different spaces.}
\label{fig:triangle}
\end{figure}
==== END The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//figures/triangle.tex ====
==== BEGIN The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//main.tex ====
\input{preamble}
\title{The maximal triangular Hilbert transform\\at the symmetric point}
\author{OpenAI}
\date{September 24, 2026}
\begin{document}
\maketitle
\begin{abstract}
For arbitrary complex inputs in $L^3(\R^2)$, we prove the pointwise
maximal $L^3\times L^3\to L^{3/2}$ estimate for the triangular Hilbert
transform, with the supremum over both hard truncation endpoints.
The estimate yields joint almost-everywhere and $L^{3/2}$ convergence
as the lower endpoint tends to zero and the upper endpoint tends to infinity.
This also proves the conjectured scalar estimate at the symmetric point.
\end{abstract}
\tableofcontents
\input{sections/introduction}
\input{sections/matrix}
\input{sections/trace}
\input{sections/heat}
\input{sections/jumps}
\input{sections/maximal}
\input{sections/consequences}
\bibliographystyle{plain}
\bibliography{references}
\end{document}
==== END The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//main.tex ====
==== BEGIN The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//preamble.tex ====
\documentclass[11pt]{article}
\pdfinfoomitdate=1
\pdftrailerid{}
\usepackage[T1]{fontenc}
\usepackage[utf8]{inputenc}
\usepackage{lmodern}
\ifdefined\pdfglyphtounicode
\pdfgentounicode=1
\pdfglyphtounicode{tfm:lmex10/parenleftbig}{0028}
\pdfglyphtounicode{tfm:lmex10/parenrightbig}{0029}
\pdfglyphtounicode{tfm:lmex10/bracketleftbig}{005B}
\pdfglyphtounicode{tfm:lmex10/bracketrightbig}{005D}
\pdfglyphtounicode{tfm:lmex10/angbracketleftbig}{27E8}
\pdfglyphtounicode{tfm:lmex10/angbracketrightbig}{27E9}
\pdfglyphtounicode{tfm:lmex10/vextendsingle}{23D0}
\pdfglyphtounicode{tfm:lmex10/vextenddouble}{2016}
\pdfglyphtounicode{tfm:lmex10/parenleftbigg}{0028}
\pdfglyphtounicode{tfm:lmex10/parenrightbigg}{0029}
\pdfglyphtounicode{tfm:lmex10/parenleftBigg}{0028}
\pdfglyphtounicode{tfm:lmex10/parenrightBigg}{0029}
\pdfglyphtounicode{tfm:lmex10/braceleftBigg}{007B}
\pdfglyphtounicode{tfm:lmex10/parenlefttp}{239B}
\pdfglyphtounicode{tfm:lmex10/parenrighttp}{239E}
\pdfglyphtounicode{tfm:lmex10/bracketlefttp}{23A1}
\pdfglyphtounicode{tfm:lmex10/bracketrighttp}{23A4}
\pdfglyphtounicode{tfm:lmex10/bracketleftbt}{23A3}
\pdfglyphtounicode{tfm:lmex10/bracketrightbt}{23A6}
\pdfglyphtounicode{tfm:lmex10/parenleftbt}{239D}
\pdfglyphtounicode{tfm:lmex10/parenrightbt}{23A0}
\pdfglyphtounicode{tfm:lmex10/parenleftex}{239C}
\pdfglyphtounicode{tfm:lmex10/parenrightex}{239F}
\pdfglyphtounicode{tfm:lmex10/summationtext}{2211}
\pdfglyphtounicode{tfm:lmex10/producttext}{220F}
\pdfglyphtounicode{tfm:lmex10/integraltext}{222B}
\pdfglyphtounicode{tfm:lmex10/summationdisplay}{2211}
\pdfglyphtounicode{tfm:lmex10/productdisplay}{220F}
\pdfglyphtounicode{tfm:lmex10/integraldisplay}{222B}
\pdfglyphtounicode{tfm:lmex10/bracketleftBig}{005B}
\pdfglyphtounicode{tfm:lmex10/bracketrightBig}{005D}
\pdfglyphtounicode{tfm:lmex10/radicalbig}{221A}
\pdfglyphtounicode{tfm:lmex10/radicalBig}{221A}
\pdfglyphtounicode{tfm:lmex10/radicalBigg}{221A}
\pdfglyphtounicode{tfm:lmsy10/mapsto}{007C}
\pdfglyphtounicode{tfm:lmsy10/B}{212C}
\pdfglyphtounicode{tfm:lmsy10/E}{2130}
\pdfglyphtounicode{tfm:lmsy10/L}{2112}
\pdfglyphtounicode{tfm:lmsy10/S}{1D4AE}
\pdfglyphtounicode{tfm:lmsy10/U}{1D4B0}
\pdfglyphtounicode{tfm:msbm10/C}{2102}
\pdfglyphtounicode{tfm:msbm10/E}{1D53C}
\pdfglyphtounicode{tfm:msbm10/R}{211D}
\pdfglyphtounicode{tfm:msbm10/Z}{2124}
\pdfglyphtounicode{tfm:lmmi10/mu}{03BC}
\pdfglyphtounicode{tfm:rm-lmr10/Delta}{0394}
\pdfglyphtounicode{tfm:rm-lmr10/Omega}{03A9}
\pdfglyphtounicode{tfm:rm-lmss10/L}{1D5AB}
\pdfglyphtounicode{tfm:rm-lmss10/R}{1D5B1}
\pdfglyphtounicode{tfm:lmmib10/alpha}{1D736}
\pdfglyphtounicode{tfm:rm-lmbx10/one}{1D7CF}
\fi
\usepackage[a4paper,margin=28mm]{geometry}
\usepackage{microtype}
\usepackage{amsmath,amssymb,amsthm,mathtools}
\usepackage{enumitem,graphicx,booktabs,tikz}
\usepackage{flafter}
\usepackage{needspace}
\usetikzlibrary{arrows.meta,positioning,calc}
\usepackage{xcolor}
\usepackage{hyperref}
\hypersetup{colorlinks=true,linkcolor=blue!45!black,citecolor=blue!45!black,
urlcolor=blue!45!black,pdftitle={The maximal triangular Hilbert transform at the symmetric point},pdfauthor={OpenAI}}
\usepackage[capitalise,noabbrev]{cleveref}
\newcommand{\R}{\mathbb R}
\newcommand{\C}{\mathbb C}
\newcommand{\dd}{\mathop{}\!\mathrm d}
\newcommand{\op}{\mathrm{op}}
\newcommand{\HS}{\mathrm{HS}}
\DeclareMathOperator{\tr}{tr}
\DeclareMathOperator{\supp}{supp}
\DeclareMathOperator{\diag}{diag}
\newtheorem{theorem}{Theorem}[section]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{corollary}[theorem]{Corollary}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\theoremstyle{remark}
\newtheorem{remark}[theorem]{Remark}
\numberwithin{equation}{section}
\setlist[enumerate]{itemsep=3pt,topsep=5pt}
\setlist[itemize]{itemsep=3pt,topsep=5pt}
\setcounter{tocdepth}{2}
==== END The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//preamble.tex ====
==== BEGIN The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/consequences.tex ====
\section{General inputs and principal values}\label{sec:consequences}
We now pass the hard maximal estimate from compactly supported smooth
inputs to all complex $L^3$ inputs. The main point is to control all
truncations on one set of full measure. We then prove joint principal-value
convergence and recover the scalar simplex formulation.
Write $p=3/2$ and
\[
B_{\varepsilon,R}(F,G)(x,y)
=\int_{\varepsilon<|t|<R}F(x+t,y)G(x,y+t)\,\frac{\dd t}{t}.
\]
Proposition~\ref{prop:compact-maximal} supplies the hard maximal estimate
for compactly supported smooth inputs; denote its constant by $C_M$.
\begin{lemma}[Finite truncations]\label{lem:finite-truncations}
For complex $F,G\in L^3(\R^2)$, there is a set of full measure on which
every finite truncation is absolutely convergent and depends continuously
on $(\varepsilon,R)$ in $0<\varepsilon<R<\infty$.
The resulting family is independent of the representatives of $F,G$ up
to a common null set. Moreover,
\begin{equation}\label{eq:elementary-bilinear}
\left\|\int_{\varepsilon<|t|<R}
|F(x+t,y)G(x,y+t)|\,\frac{\dd t}{|t|}\right\|_p
\le 2\log(R/\varepsilon)\|F\|_3\|G\|_3.
\end{equation}
\end{lemma}
\begin{proof}
For each fixed $t$, H\"older's inequality and translation invariance give
$\|F(\cdot+t,\cdot)G(\cdot,\cdot+t)\|_p\le\|F\|_3\|G\|_3$.
Minkowski's integral inequality proves
\eqref{eq:elementary-bilinear}. Apply it to the countable annuli
$1/n<|t|<n$, $n\ge2$. Outside their combined exceptional null set,
the integrand is absolutely integrable on every compact annulus away
from zero. Absolute continuity of its integral gives the asserted
endpoint continuity. If either input changes on a null set, apply the
same bound to the zero $L^3$ difference, and use the same countable
annuli. This proves independence simultaneously at every endpoint pair.
\end{proof}
Define $B_*(F,G)$ on this full-measure set as the supremum of
$|B_{\varepsilon,R}(F,G)|$ over all finite endpoint pairs, and set it to
zero elsewhere. Endpoint continuity makes this supremum equal to the
one over positive rational endpoint pairs, so it is measurable.
\begin{proof}[Proof of Theorem~\ref{thm:main}]
Choose compactly supported smooth $f_m\to F$ and $g_m\to G$ in $L^3$.
For a fixed endpoint pair, bilinearity and
\eqref{eq:elementary-bilinear} show that
$B_{\varepsilon,R}(f_m,g_m)\to B_{\varepsilon,R}(F,G)$ in $L^p$.
Fix a finite family of rational endpoint pairs. The difference of the
two finite maxima of magnitudes is bounded by the sum of the corresponding
truncation differences. Thus these maxima also converge in $L^p$.
Passing the compact smooth maximal estimate to the limit bounds this
finite maximum by $C_M\|F\|_3\|G\|_3$. The constant is independent of
the family. Exhaust all rational endpoint pairs and use monotone
convergence and Lemma~\ref{lem:finite-truncations} to obtain
Theorem~\ref{thm:main}.
We also obtain a useful continuity statement. For two pairs of inputs,
bilinearity of every truncation gives, almost everywhere,
\begin{equation}\label{eq:maximal-continuity}
|B_*(F,G)-B_*(f,g)|
\le B_*(F-f,G)+B_*(f,G-g).
\end{equation}
The bound just established makes the maximal map continuous from
$L^3\times L^3$ to $L^{3/2}$. Density shows that it is the unique
continuous extension of its compact smooth values.
\end{proof}
\begin{theorem}[Joint bilinear principal values]\label{thm:bilinear-pv}
For all complex $F,G\in L^3(\R^2)$, the joint limit
\[
B(F,G)=\lim_{\substack{\varepsilon\downarrow0\\R\uparrow\infty}}
B_{\varepsilon,R}(F,G)
\]
exists almost everywhere and in $L^{3/2}(\R^2)$, with
$\|B(F,G)\|_{3/2}\le C_M\|F\|_3\|G\|_3$.
More precisely,
\begin{equation}\label{eq:maximal-tail-convergence}
\left\|\sup_{\substack{0<\varepsilon<1/n\\R>n}}
|B_{\varepsilon,R}(F,G)-B(F,G)|\right\|_{3/2}
\longrightarrow0.
\end{equation}
The limit is the unique bounded bilinear extension of the
principal-value operator on compactly supported smooth inputs.
\end{theorem}
\begin{proof}
For compactly supported smooth $f,g$ and a fixed $(x,y)$, the function
$h(t)=f(x+t,y)g(x,y+t)$ is smooth and compactly supported. The truncated
integral is
\[
\int_\varepsilon^R\frac{h(t)-h(-t)}t\,\dd t.
\]
Its numerator is $O(t)$ at zero and vanishes for large $t$, so it has a
joint limit at every point.
For general inputs, let
\[
\Omega_n(F,G)=
\sup_{\substack{0<\varepsilon,\varepsilon'<1/n\\R,R'>n}}
|B_{\varepsilon,R}(F,G)-B_{\varepsilon',R'}(F,G)|.
\]
These are measurable: by endpoint continuity the supremum may be taken
over rational parameters. They decrease with $n$ and are bounded by
$2B_*(F,G)$. Fix smooth approximants $f_m,g_m$ as above. Outside a
single null set for this countable collection of input pairs,
\[
\Omega_n(F,G)\le\Omega_n(f_m,g_m)
+2B_*(F-f_m,G)+2B_*(f_m,G-g_m).
\]
Letting $n\to\infty$, the smooth joint convergence and
Theorem~\ref{thm:main} give
\[
\left\|\lim_{n\to\infty}\Omega_n(F,G)\right\|_p
\le 2C_M\bigl(\|F-f_m\|_3\|G\|_3
+\|f_m\|_3\|G-g_m\|_3\bigr).
\]
The right side tends to zero with $m$. Hence the limiting oscillation
is zero almost everywhere. This is the Cauchy criterion for all pairs
of endpoints in the indicated tails, so the joint complex limit exists.
It is measurable by taking a fixed cofinal sequence of endpoint pairs,
and its magnitude is at most $B_*(F,G)$.
Taking $(\varepsilon',R')$ to the limit in the definition of
$\Omega_n$ bounds the supremum in
\eqref{eq:maximal-tail-convergence} by $\Omega_n(F,G)$.
Dominated convergence applies because
$0\le\Omega_n(F,G)\le2B_*(F,G)\in L^p$.
This proves \eqref{eq:maximal-tail-convergence}, and hence joint norm
convergence. Passing to the norm limit in the bilinear identities for
the truncations proves bilinearity of $B$. Its bound and uniqueness
follow from Theorem~\ref{thm:main} and density.
\end{proof}
\subsection{Scalar forms and the simplex coordinates}
For complex $F_0,F_1,F_2\in L^3(\R^2)$, define the flat scalar form
\[
\mathcal L_{\varepsilon,R}(F_0,F_1,F_2)
=\iiint_{\varepsilon<|t|<R}
F_0(x,y)F_1(x+t,y)F_2(x,y+t)
\,\dd x\,\dd y\,\frac{\dd t}{t}.
\]
\begin{corollary}[Flat scalar principal values]\label{cor:scalar-pv}
For complex $F_0,F_1,F_2\in L^3(\R^2)$, every finite flat truncation is
absolutely integrable, and
\[
\sup_{0<\varepsilon<R<\infty}
|\mathcal L_{\varepsilon,R}(F_0,F_1,F_2)|
\le C_M\prod_{v=0}^2\|F_v\|_3.
\]
Its joint scalar principal value exists and equals
$\int F_0 B(F_1,F_2)$. This limiting form is the unique bounded
trilinear extension of its compact smooth definition.
\end{corollary}
\begin{proof}
H\"older's inequality in $(x,y)$ bounds the absolute integral by
$2\log(R/\varepsilon)\prod_v\|F_v\|_3$.
Thus Fubini identifies the truncated form with
$\int F_0 B_{\varepsilon,R}(F_1,F_2)$.
Theorem~\ref{thm:main} gives its endpoint-uniform bound, and
Theorem~\ref{thm:bilinear-pv} gives joint scalar convergence by pairing
the $L^{3/2}$ limit with $F_0$. The limiting form is bounded and
complex trilinear. Density of compact smooth triples proves uniqueness.
\end{proof}
For the simplex form, set
\[
\Lambda_{\varepsilon,R}(G_0,G_1,G_2)
=\iiint_{\varepsilon<|a+b+c|<R}
G_0(a,b)G_1(b,c)G_2(c,a)
\,\frac{\dd a\,\dd b\,\dd c}{a+b+c}.
\]
The flat and simplex scalar formulations, including equality of their
optimal norm constants, are already related in
\cite[Appendix~B.1]{KovacThieleZorinKranich2015}. We record explicit
norm-preserving maps and track both finite endpoints and their joint limit.
\begin{proposition}[Scalar equivalence]\label{prop:flat-equivalence}
For all complex $G_0,G_1,G_2\in L^3(\R^2)$, every finite simplex
truncation is absolutely integrable, and
\[
\sup_{0<\varepsilon<R<\infty}
|\Lambda_{\varepsilon,R}(G_0,G_1,G_2)|
\le C_M\prod_{v=0}^2\|G_v\|_3.
\]
Its joint scalar principal value exists for every such triple and is
the unique bounded trilinear extension of its compact smooth definition.
The optimal constants in the uniform flat and simplex scalar estimates
are equal.
\end{proposition}
\begin{proof}
Use the determinant-one change of variables
\[
x=-a,\qquad y=-b,\qquad t=a+b+c,
\qquad
a=-x,\quad b=-y,\quad c=t+x+y.
\]
For flat inputs define
\[
\begin{aligned}
G_0(u,v)&=F_0(-u,-v),\\
G_1(u,v)&=F_1(u+v,-u),\\
G_2(u,v)&=F_2(-v,u+v).
\end{aligned}
\]
Each pair map has determinant one. Their inverses are
\[
F_0(r,s)=G_0(-r,-s),\quad
F_1(r,s)=G_1(-s,r+s),\quad
F_2(r,s)=G_2(r+s,-r).
\]
They preserve $L^3$ norms, compact smoothness, and Schwartz class.
Direct substitution identifies the flat and simplex integrands and
their truncation regions. The absolute integrability, uniform bound,
and joint scalar convergence therefore follow from the flat conclusions.
The inverse maps realize every simplex triple, which proves equality
of the two optimal scalar constants. Density gives uniqueness of the
bounded trilinear extension.
\end{proof}
Write $C_S$ for the common optimal scalar constant. Complex $L^p$
duality shows that it is also the optimal constant for the uniform
family of individual bounds
\[
\|B_{\varepsilon,R}(F,G)\|_{3/2}
\le C_S\|F\|_3\|G\|_3.
\]
Indeed, for nonzero $U\in L^p$, the function
\[
H=\frac{\overline U}{|U|}\,
\frac{|U|^{1/2}}{\|U\|_p^{1/2}},
\]
set to zero where $U=0$, has $L^3$ norm one and satisfies
$\int HU=\|U\|_p$. Apply this to each bilinear truncation; the zero
case is immediate. The pointwise maximal optimal constant is at least
$C_S$.
==== END The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/consequences.tex ====
==== BEGIN The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/heat.tex ====
\section{Gaussian heat dissipation}\label{sec:heat}
We now weight the three edge matrices by Gaussians. For the variance rates
chosen below, the sum of the star energies decreases after integration over
a plane of centers. This decrease controls the matrix derivatives that will
produce the odd kernel. The estimate must be
independent of the number of matrix entries; this uniformity will allow us
to pass from finite matrices to functions.
Continuous Gaussian telescoping for entangled forms appears in
Durcik~\cite[Lemma~3 and Section~3]{Durcik2015}. Here we use the Hessian of a matrix
energy and compare its costs at neighboring vertices.
Fix $h>0$, an integer $N\ge0$, and nodes $u_i=ih$, $-N\le i\le N$.
At vertex $v\in\mathbb Z/3\mathbb Z$ take $H_v=\C^{2N+1}$, and prescribe
arbitrary complex edge values $A_v(i,j)$. For centers $p_v\in\R$ and
variances $\sigma_v>0$, define
\begin{gather*}
g_\sigma(t)=(2\pi\sigma)^{-1/2}e^{-t^2/(2\sigma)},\qquad
w_v(i)=h g_{\sigma_v}(u_i-p_v),\\
W_v(i,j)=A_v(i,j)\sqrt{w_v(i)w_{v+1}(j)}.
\end{gather*}
Use the stars $R_v=[W_v\ W_{v-1}^*]$ and Grams
$T_v=R_v^*R_v$, $S_v=R_vR_v^*$ from Section~\ref{sec:trace}, and set
\[
e_v=\Phi(T_v)=\Phi(S_v).
\]
Define the diagonal score and its insertions into the incident stars by
\begin{align*}
L_w&=\diag_i\frac{u_i-p_w}{\sigma_w},
&U_v&=R_v^*L_vR_v,\\
U_{v,v+1}&=W_vL_{v+1}W_v^*,
&U_{v,v-1}&=W_{v-1}^*L_{v-1}W_{v-1}.
\end{align*}
Define their quadratic costs and the mixed term by
\begin{equation}\label{eq:heat-costs}
V_v=b_{T_v}(U_v),\qquad
Y_{v,w}=b_{S_v}(U_{v,w})\quad(w\ne v),\qquad
Z_v=\mathcal B_{S_v}(U_{v,v+1},U_{v,v-1}).
\end{equation}
Every insertion is supported on the Gram used in its definition: an
inserted Gram $A^*BA$ annihilates $\ker A$ on both sides.
For $d\in\R$, let $\Pi_d=\{p_0+p_1+p_2=d\}$, with the measure
\begin{equation}\label{eq:heat-plane-measure}
\int_{\Pi_d}H\,\dd\pi_d
=\int_{\R^2}H(p_0,p_1,d-p_0-p_1)\,\dd p_0\,\dd p_1.
\end{equation}
Any other pair of free coordinates gives the same measure, since the
change of coordinates has absolute determinant one. Put $\lambda=11/10$.
We also write
\[
\mathcal S(A)=\sum_{v=0}^2\sum_{i,j=-N}^Nh^2|A_v(i,j)|^3.
\]
Our goal in this section is the fixed-array inequality
\[
-\partial_sJ(d,s)\ge c_*\int_{\Pi_d}\sum_vV_v\,\dd\pi_d,
\qquad J(d,s)=\int_{\Pi_d}\sum_ve_v\,\dd\pi_d,
\]
when the variances are $\sigma_i=\alpha_i s$ and
$\boldsymbol\alpha$ is a permutation of $(1,\lambda,\lambda)$.
The odd-window estimate in Section~\ref{sec:maximal} uses wider Gaussian
weights for uniformly bounded diagonal factors, but compares each score cost
after narrowing only its own vertex. The fixed enlargement
$\lambda=11/10$ retains strict dissipation for all three comparisons.
Lemma~\ref{lem:plane-dissipation} gives an absolute $c_*>0$;
Lemma~\ref{lem:initial-energy} bounds $J$ by $C\mathcal S(A)$ at scales
$s\ge h^2$. Later, when entries change at prescribed scales, the same
differential inequality will apply between successive changes.
We first differentiate the star energies, choosing the Gram orientation
whose support stays fixed. We then integrate over the center plane and
compare neighboring stars. These two steps produce the strict decrease
that controls the score costs.
\subsection{Differentiating the star energies}
\begin{lemma}\label{lem:star-smoothness}
For fixed finite edge arrays, each $e_v$ is smooth in all centers and
positive variances. If only the parameters at vertex $v$ vary, the support
of $T_v$ is fixed. If the parameters at either or both other vertices
vary while those at $v$ are fixed, the support of $S_v$ is fixed.
\end{lemma}
\begin{proof}
The weighted star has the form $R_v=D_{\rm row}R_v^0D_{\rm col}$, where
$R_v^0=[A_v\ A_{v-1}^*]$ is fixed and both diagonal multipliers are strictly
positive; the column multiplier has one block for each neighbor.
Invertible row multiplication preserves the kernel, and invertible column
multiplication preserves the range. This proves the two support
assertions, since
\[
\ker R_v=\ker(R_v^0D_{\rm col}),\qquad
\operatorname{ran}R_v=D_{\rm row}\operatorname{ran}R_v^0
\]
when the respective other multiplier is held fixed.
For joint smoothness, the rank-zero case is immediate. Otherwise factor
$R_v^0=XY^*$ with $X,Y$ of full column rank. Write
$X'=D_{\rm row}X$, $Y'=D_{\rm col}Y$ and
$P'=X'^*X'$, $Q'=Y'^*Y'$. These last two matrices are positive definite
and depend smoothly on the parameters. The nonzero eigenvalues of
$R_vR_v^*=X'Q'X'^*$ are those of $Q'^{1/2}P'Q'^{1/2}$, by comparing
$AA^*$ and $A^*A$ for $A=X'Q'^{1/2}$. Therefore
\[
e_v=\Phi(Q'^{1/2}P'Q'^{1/2}),
\]
which is smooth on the positive definite cone.
\end{proof}
\begin{lemma}\label{lem:heat-identities}
Fix a positive vector $\boldsymbol\alpha$ and set $\sigma_i=\alpha_i s$.
For all centers and all $s>0$,
\begin{align}
2\partial_s e_v
&=\sum_i\alpha_i\partial_{p_i}^2e_v
-\alpha_vV_v-\sum_{w\ne v}\alpha_wY_{v,w},
\label{eq:heat-identity}\\
\partial_{p_{v+1}}\partial_{p_{v-1}}e_v&=Z_v.
\label{eq:heat-mixed}
\end{align}
\end{lemma}
\begin{proof}
For $w=h g_\sigma(u-p)$ and $L=(u-p)/\sigma$,
\[
\partial_pw=Lw,\qquad
\partial_p^2w=(L^2-\sigma^{-1})w,\qquad
\partial_\sigma w=\tfrac12\partial_p^2w.
\]
With the columns fixed, $T_v$ is linear in the row density, so
\[
\partial_{p_v}T_v=U_v,\qquad
\partial_{\sigma_v}T_v=\tfrac12\partial_{p_v}^2T_v.
\]
Its support is fixed by Lemma~\ref{lem:star-smoothness}. The chain rule
and Lemma~\ref{lem:hessian} thus give
\[
\partial_{\sigma_v}e_v
=\tfrac12\partial_{p_v}^2e_v
-\tfrac12\mathcal B_{T_v}(U_v,U_v)
=\tfrac12\partial_{p_v}^2e_v-\tfrac12V_v.
\]
For a column vertex $w\ne v$, use instead $S_v$, which is linear in that
column density and has fixed support. This gives
\[
\partial_{p_w}S_v=U_{v,w},\qquad
\partial_{\sigma_w}e_v=\tfrac12\partial_{p_w}^2e_v-\tfrac12Y_{v,w}.
\]
Summing the three variance derivatives with coefficients $\alpha_i$
proves \eqref{eq:heat-identity}.
The two summands in $S_v=W_vW_v^*+W_{v-1}^*W_{v-1}$ depend on different
column densities. Its mixed derivative in the two column centers is
therefore zero. Only the Hessian term remains in the mixed derivative of
$e_v$, giving \eqref{eq:heat-mixed}.
\end{proof}
\subsection{Comparing neighbors and integrating centers}
\begin{lemma}\label{lem:neighbors}
For every choice of the weighted matrices and every vertex $w$,
\begin{equation}\label{eq:heat-neighbors}
Y_{w+1,w}+Y_{w-1,w}\le\sqrt2\,V_w.
\end{equation}
\end{lemma}
\begin{proof}
Abbreviate $A=W_w$, $B=W_{w-1}^*$, and $L=L_w$. The inequality
$\lVert Az+Bz'\rVert^2\le2\lVert Az\rVert^2+2\lVert Bz'\rVert^2$
implies
\[
T_w\le2Q,\qquad Q=\diag(A^*A,B^*B).
\]
Since $U_w=[A\ B]^*L[A\ B]$ is supported on $T_w$,
Lemma~\ref{lem:order} gives $b_{T_w}(U_w)\ge2^{-1/2}b_Q(U_w)$.
In a block-preserving eigenbasis of $Q$, the formula for $b_Q$ is a sum
of nonnegative squared entries. Discarding the off-diagonal blocks yields
\[
b_Q(U_w)\ge b_{A^*A}(A^*LA)+b_{B^*B}(B^*LB).
\]
Here $A^*A\le S_{w+1}$ and $B^*B\le S_{w-1}$, and the respective
insertions are $U_{w+1,w}$ and $U_{w-1,w}$. Each is supported on its
smaller edge Gram. Apply Lemma~\ref{lem:order} again, now with comparison
factor one, to obtain \eqref{eq:heat-neighbors}.
\end{proof}
\begin{lemma}\label{lem:plane-dissipation}
Fix finite edge arrays and a positive vector $\boldsymbol\alpha$, and
evaluate the energies at $\sigma_i=\alpha_i s$. Set
\[
J_v(d,s)=\int_{\Pi_d}e_v\,\dd\pi_d,
\qquad J(d,s)=\sum_vJ_v(d,s),
\qquad \Sigma_\alpha=\sum_i\alpha_i.
\]
These integrals are finite, continuously differentiable in $s>0$, and
twice continuously differentiable in $d$. Moreover,
\begin{equation}\label{eq:heat-plane-dissipation}
2\partial_sJ(d,s)\le\int_{\Pi_d}\sum_w
\left[-\alpha_w+
\sqrt2\max\{0,\Sigma_\alpha/2-\alpha_w\}\right]V_w\,\dd\pi_d.
\end{equation}
In particular, for each permutation of $(1,\lambda,\lambda)$,
\begin{equation}\label{eq:heat-strict}
-\partial_sJ(d,s)\ge c_*\int_{\Pi_d}\sum_wV_w\,\dd\pi_d,
\qquad
c_* = \frac12\min\{1-\tfrac35\sqrt2,\tfrac{11}{10}-\tfrac12\sqrt2\}>0.
\end{equation}
\end{lemma}
\begin{proof}
We first justify differentiation at the fixed mesh. On compact positive
$s$-intervals the nodes are bounded, the scores grow at most linearly in
the centers, and the first and second density ratios grow at most
quadratically. Lemma~\ref{lem:insertion} gives
\[
V_v\le\tfrac34\lVert L_v\rVert_{\op}^2e_v,
\qquad Y_{v,w}\le\tfrac34\lVert L_w\rVert_{\op}^2e_v.
\]
For the second inequality, pad $L_w$ by zero on the other column block,
obtaining $B_w$ with $U_{v,w}=R_vB_wR_v^*$, and apply the lemma to
$R_v^*$. Also $2|Z_v|\le Y_{v,v+1}+Y_{v,v-1}$.
For a Hermitian matrix $B$,
\[
|\tr(\sqrt{T_v}R_v^*BR_v)|\le\lVert B\rVert_{\op}e_v,
\]
since $R_v\sqrt{T_v}R_v^*$ is positive with trace $e_v$. The corresponding
column estimate follows in the same way with $S_v$ and a padded column
insertion. Together with the density-ratio formulas in
Lemma~\ref{lem:heat-identities}, these bounds control every first center
derivative, pure second center derivative, mixed column derivative, and
first $s$-derivative by a polynomial in the centers times $e_v$.
To dominate the energy itself, use
\begin{equation}\label{eq:heat-energy-bound}
e_v\le(\lVert W_v\rVert_{\HS}^2+
\lVert W_{v-1}\rVert_{\HS}^2)^{3/2}
\le\sqrt2(\lVert W_v\rVert_{\HS}^3+
\lVert W_{v-1}\rVert_{\HS}^3).
\end{equation}
Each edge norm cubed is bounded by a Gaussian in its two endpoint
centers, locally uniformly for $s>0$. Indeed, the finite set of nodes is
bounded and $(u-p)^2\ge p^2/2-u^2$. On $\Pi_d$ with $|d|\le D$, any
pair of centers satisfies
\[
p_0^2+p_1^2+p_2^2\le4(p_i^2+p_j^2)+3D^2.
\]
Thus an endpoint-pair Gaussian, multiplied by any of the preceding
polynomials, has an integrable majorant in any fixed plane chart, locally
uniformly in $(d,s)$. This proves all needed differentiation statements
by dominated convergence. The constants in this justification may
depend on the finite mesh and the compact parameter intervals; they will
not enter the uniform estimate.
We can now differentiate the plane integrals. It remains to express
their second derivatives in $d$ using both pure and mixed center
derivatives, then apply Lemma~\ref{lem:neighbors} to the resulting costs.
Choosing $p_i$ as the dependent coordinate in
\eqref{eq:heat-plane-measure} gives
\[
\partial_d^2J_v=\int_{\Pi_d}\partial_{p_i}^2e_v\,\dd\pi_d
\qquad\text{for every }i.
\]
For the column indices $j=v+1$, $k=v-1$, first choose $p_j$ dependent to
write $\partial_dJ_v=\int_{\Pi_d}\partial_{p_j}e_v\,\dd\pi_d$.
Reparametrize this last integral with $p_k$ dependent, then differentiate
again. It follows from \eqref{eq:heat-mixed} that
\begin{equation}\label{eq:heat-plane-mixed}
\partial_d^2J_v
=\int_{\Pi_d}\partial_{p_k}\partial_{p_j}e_v\,\dd\pi_d
=\int_{\Pi_d}Z_v\,\dd\pi_d.
\end{equation}
These are identities of integrated derivatives; no pointwise equality
between pure and mixed derivatives is required.
Integrating \eqref{eq:heat-identity} and using
\eqref{eq:heat-plane-mixed}, we obtain
\[
2\partial_sJ_v=\int_{\Pi_d}
\left[\Sigma_\alpha Z_v-\alpha_vV_v
-\sum_{w\ne v}\alpha_wY_{v,w}\right]\dd\pi_d.
\]
Cauchy--Schwarz gives $2Z_v\le Y_{v,v+1}+Y_{v,v-1}$. Sum over $v$ and
collect the two neighbor costs carrying the score at $w$. Their common
coefficient is $\Sigma_\alpha/2-\alpha_w$. If it is negative, discard
those terms; otherwise apply Lemma~\ref{lem:neighbors}. This proves
\eqref{eq:heat-plane-dissipation}.
For the indicated permutations, $\Sigma_\alpha=16/5$. The two possible
coefficients in \eqref{eq:heat-plane-dissipation} are
$-1+(3/5)\sqrt2$ and $-11/10+(1/2)\sqrt2$, both strictly negative. Dividing by two gives
\eqref{eq:heat-strict}.
\end{proof}
\subsection{The initial energy}
The previous lemma controls the accumulated costs by the energy lost along
the heat flow. We now bound the energy itself at scales no smaller than the
mesh. This estimate will also apply immediately after entries are switched
on or off, since it requires only a bound on their absolute values.
\begin{lemma}[Initial energy]\label{lem:initial-energy}
Let $h>0$, let $N\ge0$ be an integer, and let $A_v(i,j)$ be the finite complex arrays above.
For any $d\in\R$ and variances $\sigma_v\ge h^2$,
\begin{equation}\label{eq:initial-energy}
\int_{\Pi_d}\sum_ve_v\,\dd\pi_d
\le C_0\mathcal S(A),\qquad
C_0=2\sqrt2\left(1+\sqrt{2/\pi}\right).
\end{equation}
The same bound holds if any entries are replaced by smaller entries in
absolute value before forming the weighted matrices.
\end{lemma}
\begin{proof}
For every center $p$ and variance $\sigma>0$,
\begin{equation}\label{eq:heat-mesh-mass}
\sum_{i=-N}^Nh g_\sigma(u_i-p)
\le1+\frac{2h}{\sqrt{2\pi\sigma}}.
\end{equation}
To see this, compare a grid point $u_i\ge p+h$ with the integral over its
preceding interval of length $h$, where the Gaussian is at least its
value at $u_i$. For $u_i\le p-h$, use the succeeding interval. These
intervals are disjoint, and the at most two remaining grid points each
contribute at most $h\sup g_\sigma$.
Since each $\sigma_v\ge h^2$, every density has total mass at most
$M=1+\sqrt{2/\pi}$. Weighted H\"older gives
\begin{align*}
\lVert W_v\rVert_{\HS}^3
&=\left(\sum_{i,j}|A_v(i,j)|^2w_v(i)w_{v+1}(j)\right)^{3/2}\\
&\le M\sum_{i,j}|A_v(i,j)|^3w_v(i)w_{v+1}(j).
\end{align*}
Integrate with $p_v,p_{v+1}$ as free coordinates on $\Pi_d$. Each density
has integral $h$, so
\[
\int_{\Pi_d}\lVert W_v\rVert_{\HS}^3\,\dd\pi_d
\le M\sum_{i,j}h^2|A_v(i,j)|^3.
\]
Each edge occurs in exactly two stars. Equation~\eqref{eq:heat-energy-bound}
therefore proves \eqref{eq:initial-energy}. Replacing entries by smaller
ones in absolute value only reduces the last upper bounds, proving the
final assertion.
\end{proof}
For fixed arrays and a permutation $\boldsymbol\alpha$ of
$(1,\lambda,\lambda)$, Lemmas~\ref{lem:plane-dissipation} and
\ref{lem:initial-energy} now give, whenever $h^2\le s_0<s_1$,
\[
\int_{s_0}^{s_1}\int_{\Pi_0}\sum_vV_v\,\dd\pi_0\,\dd s
\le\frac{J(0,s_0)-J(0,s_1)}{c_*}
\le\frac{C_0}{c_*}\mathcal S(A).
\]
The first inequality retains the final energy. In the next section, we
will add these inequalities over intervals with different masked arrays.
==== END The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/heat.tex ====
==== BEGIN The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/introduction.tex ====
\section{Introduction}\label{sec:introduction}
The triangular Hilbert transform couples translations in two different
coordinate directions. For complex functions $F,G$ on $\R^2$ and
$0<\varepsilon<R<\infty$, write
\[
B_{\varepsilon,R}(F,G)(x,y)
=\int_{\varepsilon<|t|<R}F(x+t,y)G(x,y+t)\,\frac{\dd t}{t},
\qquad
B_*(F,G)=\sup_{0<\varepsilon<R<\infty}|B_{\varepsilon,R}(F,G)|.
\]
For $L^3$ inputs, the finite integrals are absolutely convergent outside
one common null set, and the supremum is measurable; the details are
given in Lemma~\ref{lem:finite-truncations}.
\begin{theorem}\label{thm:main}
There is an absolute constant $C$ such that, for every pair of complex
functions $F,G\in L^3(\R^2)$,
\[
\|B_*(F,G)\|_{L^{3/2}(\R^2)}
\le C\|F\|_{L^3(\R^2)}\|G\|_{L^3(\R^2)}.
\]
The supremum is over both finite hard truncation endpoints.
\end{theorem}
The estimate controls a truncation chosen separately at each spatial
point. This is stronger than a uniform bound for the norm of each
individual truncation: the same $L^{3/2}$ majorant must control every
choice of the two endpoints. It also gives the joint principal value
\[
B(F,G)=\lim_{\substack{\varepsilon\downarrow0\\R\uparrow\infty}}
B_{\varepsilon,R}(F,G)
\]
almost everywhere and in $L^{3/2}$, with the same norm bound
(Theorem~\ref{thm:bilinear-pv}).
Pairing the bilinear output with a third $L^3$ function leads to the
scalar triangular form. In its symmetric coordinates this is
\[
\Lambda_{\varepsilon,R}(G_0,G_1,G_2)
=\iiint_{\varepsilon<|a+b+c|<R}
G_0(a,b)G_1(b,c)G_2(c,a)
\,\frac{\dd a\,\dd b\,\dd c}{a+b+c}.
\]
The three factors have equal exponent $3$, which is the symmetric point
of the scalar H\"older relation $1/p_0+1/p_1+1/p_2=1$.
We obtain an endpoint-uniform bound by $C\prod_v\|G_v\|_3$ and a
joint scalar principal value for every complex $L^3$ triple.
Proposition~\ref{prop:flat-equivalence} gives the explicit
determinant-one changes of variables and proves that the flat and
simplex scalar estimates have the same optimal constant.
In particular, the result resolves positively the symmetric triangular
Hilbert transform problem stated by Thiele
in~\cite[Section~9, Problem~13]{GrafakosEtAl2017}.
\subsection{Previous work}
The flat operator appears in Demeter and Thiele's discussion of
bilinear averages for two commuting transformations~\cite[Section~6]{DemeterThiele2010}.
The three inputs of the scalar form depend on the three pairs of
variables, making it a basic example of an entangled singular integral.
Kova\v{c}, Thiele and Zorin-Kranich proved estimates for a Walsh model
with one structurally restricted input~\cite[Theorem~1.7]{KovacThieleZorinKranich2015}.
Tao proved sublogarithmic cancellation for one-dimensional multilinear
Hilbert transforms~\cite[Theorem~1.2]{Tao2016Cancellation}.
Zorin-Kranich extended this cancellation to simplex Hilbert
transforms~\cite[Theorem~1.3]{ZorinKranich2017}. Square-function estimates of Durcik,
Kova\v{c}, \v{S}kreb and Thiele gave square-root cancellation for the
triangular form across finitely many smooth scales~\cite[Corollary~5]{DurcikKovacSkrebThiele2019}.
Durcik, Kova\v{c} and Thiele proved power-type cancellation with hard
truncations~\cite[Theorem~1]{DurcikKovacThiele2019}; at the symmetric
tuple the bound grows like $\sqrt{\log(R/\varepsilon)}$, by permutation
and interpolation of their $(4,4,2)$ estimates.
These finite-scale bounds retain a dependence on the scale range.
Bounds for averages over direction parameters were obtained by
Durcik and Roos~\cite[Theorem~1]{DurcikRoos2021} and, for rough angular
weights, by Lin and Slav\'ikov\'a~\cite[Theorem~1.1]{LinSlavikova2026}.
For curved translations, Christ, Durcik and Roos proved bounds along
$(t,t^2)$~\cite[Theorem~1]{ChristDurcikRoos2021}. Hsu and Lin treat
more general curves under geometric hypotheses excluding the flat curve
$(t,t)$~\cite[Theorem~1.37]{HsuLin2026}. The curvature in these results
provides an additional source of cancellation that is unavailable for
the flat curve.
Our argument uses a nonnegative matrix energy and Gaussian heat flow.
Energy telescoping for entangled forms appears in Kova\v{c}'s
twisted-paraproduct argument~\cite[Lemma~2]{Kovac2012Twisted} and his
Bellman-function framework~\cite[Section~2]{Kovac2011}. Continuous
Gaussian telescoping appears in Durcik~\cite[Lemma~3 and Section~3]{Durcik2015}.
These methods concern bipartite forms; Kova\v{c} discusses the obstacle
posed by the triangular cycle in~\cite[Section~6.2]{Kovac2011}.
The cyclic trace representation itself also occurs
in~\cite[Equation~(1.5)]{KovacThieleZorinKranich2015}.
We prove a dimension-independent mixed trace inequality and compare
the energy derivatives at neighboring vertices. The key additional
estimate controls the total change in energy when individual matrix
entries are switched on and off at independently chosen scales.
This permits the pointwise maximum without a factor depending on
the number of scales or the matrix dimensions.
\subsection{Proof overview}
The change of variables $x=-a$, $y=-b$, $t=a+b+c$ writes the bilinear
output in terms of
\[
G_1(b,c)=F(b+c,-b),\qquad G_2(c,a)=G(-a,c+a).
\]
Both pair maps preserve $L^3$ norms. We first sample these inputs,
together with a dual input $G_0(a,b)$, on a finite mesh. Multiplying
entries by Gaussian weights gives maps $W_v:H_{v+1}\to H_v$ between
finite-dimensional complex Hilbert spaces, indexed by
$v\in\mathbb Z/3\mathbb Z$.
The three vertex spaces correspond to the sampled $a$-, $b$- and
$c$-coordinates. Thus the two indices of the dual array record the
output position $(x,y)=(-a,-b)$.
At a vertex, combine its two incident maps into
$R_v=[\,W_v\quad W_{v-1}^*\,]$ and use the energy
$\tr((R_v^*R_v)^{3/2})$.
Sections~\ref{sec:matrix} and~\ref{sec:trace} prove the matrix
inequalities. Section~\ref{sec:heat} then integrates the sum of the
three energies over Gaussian centers satisfying $p_0+p_1+p_2=0$.
For fixed arrays, its decrease as the variance grows controls the quadratic
costs needed to estimate a cyclic trace with three odd Gaussian windows.
To select different truncations at different output points, the dual
array entry at $(i,j)$ must be present only between its own two scales.
The energy is therefore decreasing between finitely many switching
scales and may jump at those scales. Section~\ref{sec:jumps} bounds
the sum of the absolute jumps by the discrete $L^3$ masses of the three
arrays. The underlying matrix fact is that the derivative with respect
to an entry is controlled by the product of its row and column norms.
After Gaussian integration, one-dimensional maximal functions bound
these products uniformly in the switching scales.
Section~\ref{sec:maximal} uses this bound to linearize any finite
maximum of smooth truncations. Mesh limits and rational exhaustion
give the full smooth maximum. A pointwise comparison of its kernel
with the hard Hilbert kernel completes the estimate for compactly
supported smooth inputs. Finally, Section~\ref{sec:consequences}
uses density and a vanishing tail-oscillation argument to prove
Theorem~\ref{thm:main}, the joint principal values, and the scalar
equivalence.
==== END The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/introduction.tex ====
==== BEGIN The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/jumps.tex ====
\section{Output-dependent truncations and energy jumps}\label{sec:jumps}
We now allow each entry of one edge array to have its own truncation
interval. Between endpoint scales, the heat estimate from
Section~\ref{sec:heat} still applies. The new task is to bound the total
increase in energy when entries switch on or off. We will in fact bound
the sum of the absolute energy jumps. The estimate depends on the
original arrays alone, because every entry switches at most twice.
We first establish two inputs to this bound: estimates for Gaussian
averages that are uniform in the scale, and a formula for changes in
the matrix energy that remains valid when the rank changes.
\subsection{Maximal averages on a line}
For a locally integrable function on $\R$, and a sequence on $\mathbb Z$,
respectively, define the centered maximal functions
\[
M_{\R}f(x)=\sup_{r>0}\frac1{2r}\int_{x-r}^{x+r}|f(t)|\,\dd t,
\qquad
M_{\mathbb Z}f(n)=\sup_{r\in\mathbb Z_{\ge0}}
\frac1{2r+1}\sum_{|m-n|\le r}|f(m)|.
\]
Thus discrete averages include singletons. We use Lebesgue measure on
$\R$ and counting measure on $\mathbb Z$.
The estimates below are versions of the classical Hardy--Littlewood
maximal inequalities~\cite[Section~IV]{HardyLittlewood1930}; we include
the centered covering argument and the Gaussian comparison with the
constants needed here.
\Needspace{3\baselineskip}
\begin{lemma}[One-dimensional maximal estimates]\label{lem:line-maximal}
On either of these measure spaces, write $M$ for the corresponding
centered maximal operator. For $f\in L^1$ and $\rho>0$,
\[
\mu\{Mf>\rho\}\le\frac3\rho\|f\|_1.
\]
For every $1<q<\infty$ and $f\in L^q$,
\begin{equation}\label{eq:line-maximal-strong}
\|Mf\|_q^q\le
\frac{6q\,2^{q-1}}{q-1}\|f\|_q^q.
\end{equation}
Moreover, there is an absolute constant $C_g$ such that, whenever
$h>0$, $\sigma\ge h^2$ and $n\in\mathbb Z$,
\begin{equation}\label{eq:discrete-gaussian-maximal}
\sum_{m\in\mathbb Z}h g_\sigma((m-n)h)|f(m)|
\le C_g M_{\mathbb Z}f(n).
\end{equation}
\end{lemma}
\begin{proof}
For a finite collection of intervals, repeatedly choose an interval of
largest radius and discard all intervals meeting it. The chosen
intervals are disjoint. Every discarded interval has radius no larger
than that of the interval which removed it and is contained in its
triple-radius enlargement. This enlargement has three times the
original length on $\R$. On $\mathbb Z$, its cardinality is
$6r+1\le3(2r+1)$, including $r=0$.
For each point in a finite subset of $\{M_{\mathbb Z}f>\rho\}$, choose
a witnessing interval whose average exceeds $\rho$. The preceding
selection bounds the size of that subset by three times the sum of
the cardinalities of the chosen intervals, and hence by
$3\|f\|_1/\rho$. Taking the supremum over finite subsets proves the
discrete weak estimate. On $\R$, fixed-radius averages of an
$L^1$ function are continuous, so $\{M_{\R}f>\rho\}$ is open.
Apply the same selection to a finite witnessing cover of any compact
subset and then use inner regularity to obtain the continuous estimate.
For $f\in L^q$, split $f$ into its parts on $\{|f|\le\rho/2\}$ and
$\{|f|>\rho/2\}$. The first part has maximal function at most
$\rho/2$, and the second is integrable. The weak estimate gives
\[
\mu\{Mf>\rho\}
\le\frac6\rho\int_{|f|>\rho/2}|f|\,\dd\mu.
\]
The distribution formula and Tonelli's theorem now yield
\[
\begin{aligned}
\|Mf\|_q^q
&\le6q\int |f(x)|
\int_0^{2|f(x)|}\rho^{q-2}\,\dd\rho\,\dd\mu(x)\\
&=\frac{6q\,2^{q-1}}{q-1}\|f\|_q^q.
\end{aligned}
\]
These formulas are valid first with a bounded range of distribution
parameters and then by monotone convergence; finiteness of $Mf$ is
not assumed in advance. Both maximal functions are measurable: the
discrete supremum is countable, and continuous averages are continuous
in their radius, so positive rational radii suffice on $\R$.
Finally put $a=\sqrt\sigma/h\ge1$. In the sum in
\eqref{eq:discrete-gaussian-maximal}, the kernel is
\[
h g_\sigma((m-n)h)
=\frac1{\sqrt{2\pi}a}
\exp\left(-\frac{(m-n)^2}{2a^2}\right).
\]
Split the integers into $|m-n|\le a$ and the annuli
$2^{j-1}a<|m-n|\le2^ja$, $j\ge1$. The sum of $|f|$ over
$|m-n|\le2^ja$ is at most $3\cdot2^ja\,M_{\mathbb Z}f(n)$.
Bounding the Gaussian by its maximum on each region proves the claim
with, for example,
\[
C_g=\frac3{\sqrt{2\pi}}
\left(1+\sum_{j\ge1}2^j e^{-4^{j-1}/2}\right)<\infty.
\]
All constants are independent of the mesh and of the support of $f$.
\end{proof}
\subsection{Changing a matrix entry}
At a truncation endpoint the rank of a star can change. We therefore
need a first-derivative formula that does not require fixed support.
For the sum of cubes of the singular values, this formula is a case
of the general gradient formula for unitarily invariant matrix
functions~\cite[Theorem~3.1]{Lewis1995}. We give a direct regularization
proof and then establish the entrywise bound used in the jump estimate.
\begin{lemma}[A derivative through rank changes]\label{lem:rank-change}
For finite rectangular complex matrices $R_-,R_+$ of the same size,
put $R(z)=(1-z)R_-+zR_+$ and $D=R_+-R_-$. Then
\begin{equation}\label{eq:rank-change-derivative}
\begin{aligned}
&\Phi(R_+^*R_+)-\Phi(R_-^*R_-)\\
&\quad=3\int_0^1\operatorname{Re}
\langle R(z)\sqrt{R(z)^*R(z)},D\rangle_{\HS}\,\dd z.
\end{aligned}
\end{equation}
For every rectangular matrix $R$ and every entry $(i,j)$,
\begin{equation}\label{eq:entry-gradient}
|(R\sqrt{R^*R})_{ij}|
\le\|\operatorname{row}_iR\|_2\,
\|\operatorname{col}_jR\|_2.
\end{equation}
\end{lemma}
\begin{proof}
For $\eta>0$, differentiate
$\tr((R(z)^*R(z)+\eta I)^{3/2})$ using
Lemma~\ref{lem:hessian}. Its derivative is
\[
3\operatorname{Re}
\langle R(z)\sqrt{R(z)^*R(z)+\eta I},D\rangle_{\HS}.
\]
Spectral calculus gives
\[
\|\sqrt{R^*R+\eta I}-\sqrt{R^*R}\|_{\op}\le\sqrt\eta.
\]
Since $R(z)$ stays in a bounded set, the derivatives converge uniformly
in $z$, and the endpoint energies converge. Integrating the
regularized identity before letting $\eta\downarrow0$ proves
\eqref{eq:rank-change-derivative}.
Cauchy--Schwarz in the matrix-product formula bounds the left side of
\eqref{eq:entry-gradient} by the norm of row $i$ of $R$ times the norm
of column $j$ of $\sqrt{R^*R}$. The squared norm of that column is
$(R^*R)_{jj}$, which is also the squared norm of column $j$ of $R$.
This proves \eqref{eq:entry-gradient}, including zero columns and
singular matrices.
\end{proof}
\subsection{The total jump bound}
Return to the finite arrays and Gaussian weights of
Section~\ref{sec:heat}, and write $\dd\pi=\dd\pi_0$ on $\Pi_0$.
Independently for every pair $(i,j)$, choose
finite endpoints
\[
h\le\varepsilon_{ij}<\rho_{ij},
\qquad
A_0^s(i,j)=A_0(i,j)
\mathbf1_{\{\varepsilon_{ij}^2<s<\rho_{ij}^2\}}.
\]
The chosen endpoints are fixed independently of the Gaussian centers.
Keep $A_1,A_2$ fixed and form the weighted matrices and costs using
$A_0^s$ in place of $A_0$. The arrays are complex; any phase at an
output pair may be included in the fixed value $A_0(i,j)$.
Set
\[
s_- =\min_{i,j}\varepsilon_{ij}^2,
\qquad s_+=\max_{i,j}\rho_{ij}^2.
\]
\begin{proposition}[Dissipation with output-dependent endpoints]
\label{prop:masked-dissipation}
For every $h>0$, $N\ge0$, complex arrays $A_v$ on
$\{-N,\ldots,N\}^2$, and endpoint choices as above, let
$\alpha$ be any permutation of $(1,11/10,11/10)$ and use variances
$\sigma_v=\alpha_vs$. The corresponding masked costs satisfy
\begin{equation}\label{eq:masked-dissipation}
\int_{s_-}^{s_+}\int_{\Pi_0}
\sum_{v=0}^2 V_v(p,s)\,\dd\pi\,\dd s
\le C\mathcal S(A).
\end{equation}
The constant is independent of the arrays, the mesh, the endpoints,
and the number of distinct endpoint values.
\end{proposition}
\begin{proof}
Write $J(s)=\int_{\Pi_0}\sum_ve_v(p,s)\,\dd\pi$ for the masked
energy, and put $\Delta J(\tau)=J(\tau+)-J(\tau-)$ at an endpoint.
We first estimate its jump at a single scale, then sum those
estimates and telescope the heat inequality.
\paragraph{The two stars at a jump.}
At an endpoint scale $\tau$, hold the weights at $s=\tau$ and let
\[
\delta A(i,j)=A_0^{\tau+}(i,j)-A_0^{\tau-}(i,j).
\]
The superscripts denote one-sided masks. Interpolate all changed
entries at once between these two arrays. Each interpolated entry
has absolute value at most $|A_0(i,j)|$, even when some entries open
and others close at the same scale.
For fixed $i,j,p$, define nonnegative numbers using the original,
unmasked arrays:
\begin{equation}\label{eq:jump-slice-quantities}
\begin{aligned}
X^2&=\sum_m|A_0(i,m)|^2w_1(m),&
Y^2&=\sum_m|A_0(m,j)|^2w_0(m),\\
H^2&=\sum_m|A_2(m,i)|^2w_2(m),&
K^2&=\sum_m|A_1(j,m)|^2w_2(m).
\end{aligned}
\end{equation}
In $R_0=[W_0\;W_2^*]$, a changing entry has row norm at most
$\sqrt{w_0(i)}\sqrt{X^2+H^2}$, column norm at most
$\sqrt{w_1(j)}Y$, and derivative modulus
$\sqrt{w_0(i)w_1(j)}|\delta A(i,j)|$.
In $R_1=[W_1\;W_0^*]$, its adjoint entry has row norm at most
$\sqrt{w_1(j)}\sqrt{Y^2+K^2}$ and column norm at most
$\sqrt{w_0(i)}X$, with the same derivative modulus.
Star $R_2=[W_2\;W_1^*]$ is unchanged. Lemma~\ref{lem:rank-change}
and $\sqrt{x^2+y^2}\le x+y$ therefore give
\begin{equation}\label{eq:jump-entry-bound}
\begin{aligned}
|\Delta e_0|+|\Delta e_1|
\le3\sum_{i,j}|\delta A(i,j)|w_0(i)w_1(j)
\bigl[(X+H)Y+(Y+K)X\bigr].
\end{aligned}
\end{equation}
This estimate requires no assumption on the ranks during interpolation.
\paragraph{A probability measure on the center plane.}
For each fixed $i,j$, the measure
\[
\dd\nu_{ij}=h^{-2}w_0(i)w_1(j)\,\dd\pi
\]
is a probability measure. Indeed, in the free coordinates $(p_0,p_1)$,
its density is
$g_{\sigma_0}(u_i-p_0)g_{\sigma_1}(u_j-p_1)$.
Thus $p_0,p_1$ are independent Gaussian variables with means
$u_i,u_j$ and variances $\sigma_0,\sigma_1$, while
$p_2=-p_0-p_1$. Adding Gaussian variances gives
\begin{equation}\label{eq:jump-gaussian-expectations}
\begin{aligned}
\mathbb E_{\nu_{ij}}w_0(m)
&=h g_{2\sigma_0}(u_m-u_i),\\
\mathbb E_{\nu_{ij}}w_1(m)
&=h g_{2\sigma_1}(u_m-u_j),\\
\mathbb E_{\nu_{ij}}w_2(m)
&=h g_{\sigma_0+\sigma_1+\sigma_2}(u_m+u_i+u_j).
\end{aligned}
\end{equation}
For the last identity, $p_0+p_1$ has mean $u_i+u_j$ and variance
$\sigma_0+\sigma_1$, and averaging $g_{\sigma_2}(u_m+p_0+p_1)$
adds the remaining variance. Each identity also follows directly
by completing the square in its Gaussian integral.
Extend the arrays by zero to $\mathbb Z^2$ and define
\begin{equation}\label{eq:jump-starred-majorants}
\begin{aligned}
X_*^2(i,j)&=M_{\mathbb Z}(|A_0(i,\cdot)|^2)(j),\\
Y_*^2(i,j)&=M_{\mathbb Z}(|A_0(\cdot,j)|^2)(i),\\
H_*^2(i,j)&=M_{\mathbb Z}(|A_2(\cdot,i)|^2)(-i-j),\\
K_*^2(i,j)&=M_{\mathbb Z}(|A_1(j,\cdot)|^2)(-i-j).
\end{aligned}
\end{equation}
Every variance in \eqref{eq:jump-gaussian-expectations} is at least
$h^2$. Lemma~\ref{lem:line-maximal} consequently bounds
$\mathbb E X^2,\mathbb E Y^2,\mathbb E H^2,\mathbb E K^2$ by
$C_g$ times the corresponding starred squares. Cauchy--Schwarz
under the probability $\nu_{ij}$ gives, in particular,
\[
\mathbb E(XY)\le C_gX_*Y_*,\qquad
\mathbb E(HY)\le C_gH_*Y_*,\qquad
\mathbb E(KX)\le C_gK_*X_*.
\]
No independence of these four random variables is needed.
The starred quantities depend only on the original arrays and the
output indices, not on the scale $\tau$.
\paragraph{Summing the jumps.}
Integrating \eqref{eq:jump-entry-bound} over the centers and using
$w_0(i)w_1(j)\,\dd\pi=h^2\,\dd\nu_{ij}$ yields
\[
|\Delta J(\tau)|\le3C_g\sum_{i,j}h^2|\delta A(i,j)|
(2X_*Y_*+H_*Y_*+K_*X_*).
\]
Every entry opens once and closes once. Therefore its total absolute
change, summed over any subset of endpoint scales, is at most
$2|A_0(i,j)|$. We obtain
\begin{equation}\label{eq:total-jump-majorant}
\sum_\tau|\Delta J(\tau)|
\le6C_g\sum_{i,j}h^2|A_0(i,j)|
(2X_*Y_*+H_*Y_*+K_*X_*).
\end{equation}
The sum may include all endpoints or just the interior ones.
To bound this expression, set
\[
n_v=\left(\sum_{i,j}h^2|A_v(i,j)|^3\right)^{1/3},
\qquad \|f\|_{3,h}^3=\sum_{i,j\in\mathbb Z}h^2|f(i,j)|^3.
\]
The $q=3/2$ case of \eqref{eq:line-maximal-strong}, applied on slices,
gives
\begin{equation}\label{eq:jump-slice-norms}
\|X_*\|_{3,h},\ \|Y_*\|_{3,h}\le Cn_0,
\qquad \|H_*\|_{3,h}\le Cn_2,
\qquad \|K_*\|_{3,h}\le Cn_1.
\end{equation}
For $H_*$, fix $i$ and change the summation index from $j$ to $-i-j$;
this is a bijection of $\mathbb Z$. For $K_*$, fix $j$ and change
$i$ to $-i-j$. Thus both estimates are ordinary slice estimates,
with no multiplicity or mesh factor. Restricting the sums back to
the original finite square can only decrease them.
H\"older's inequality in \eqref{eq:total-jump-majorant}, followed by
\eqref{eq:jump-slice-norms}, proves
\begin{equation}\label{eq:total-jump-bound}
\sum_\tau|\Delta J(\tau)|
\le C(2n_0^3+n_0^2n_2+n_0^2n_1)
\le C\mathcal S(A).
\end{equation}
The last step uses $a^2b\le(2a^3+b^3)/3$.
In particular, the estimate has no dependence on how many distinct
scales occur.
\paragraph{Telescoping the heat estimate.}
List the distinct squared endpoints as
$t_0=s_-<t_1<\cdots<t_m=s_+$.
On every open interval $(t_k,t_{k+1})$ the array is fixed, so
Lemma~\ref{lem:plane-dissipation} applies. Its fixed-array continuity
and Gaussian domination also justify the one-sided endpoint limits.
Integrating on these intervals and summing gives
\begin{align*}
c_*\int_{s_-}^{s_+}\int_{\Pi_0}\sum_vV_v\,\dd\pi\,\dd s
&\le\sum_{k=0}^{m-1}\bigl[J(t_k+)-J(t_{k+1}-)\bigr]\\
&=J(s_-+)-J(s_+-)
+\sum_{k=1}^{m-1}\bigl[J(t_k+)-J(t_k-)\bigr]\\
&\le J(s_-+)+\sum_{k=1}^{m-1}|\Delta J(t_k)|.
\end{align*}
Here the terminal energy was discarded by nonnegativity. The initial
right-hand mask has entries bounded by those of the original array,
and all its variances are at least $h^2$. Lemma~\ref{lem:initial-energy}
therefore gives $J(s_-+)\le C\mathcal S(A)$.
Together with \eqref{eq:total-jump-bound}, this proves
\eqref{eq:masked-dissipation}. The initial opening is accounted for
by the right-hand initial energy; a final closing jump is unnecessary.
Simultaneous endpoints have already been combined in the list of
distinct scales, and their values at the single endpoint do not affect
the scale integral.
\end{proof}
Figure~\ref{fig:mask-energy} illustrates the masks and the energy accounting.
\input{figures/masks}
==== END The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/jumps.tex ====
==== BEGIN The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/matrix.tex ====
\section{A matrix energy and its Hessian}\label{sec:matrix}
We begin with the finite-dimensional inequalities that control the
derivatives of our energy. Their constants are independent of the
matrix dimensions. Throughout this section, matrices act on
finite-dimensional complex Hilbert spaces, and inequalities between
Hermitian matrices mean inequalities of quadratic forms.
For a positive semidefinite matrix $P$, set
\[
\Phi(P)=\tr(P^{3/2}),
\qquad \supp P=(\ker P)^\perp.
\]
A matrix is \emph{supported on $P$} if it vanishes on $\ker P$ on
both sides. We use the Hilbert--Schmidt inner product
$\langle K,L\rangle_{\HS}=\tr(K^*L)$.
For a Hermitian matrix $A$, write
$\mathsf L_A(X)=AX$ and $\mathsf R_A(X)=XA$.
On matrices acting on $\supp P$, the operator
$\mathsf L_{\sqrt P}+\mathsf R_{\sqrt P}$ is positive definite.
For Hermitian $K,L$ supported on $P$, define
\[
\mathcal B_P(K,L)
=\frac32\operatorname{Re}
\big\langle K,
(\mathsf L_{\sqrt P}+\mathsf R_{\sqrt P})^{-1}L
\big\rangle_{\HS},
\qquad b_P(K)=\mathcal B_P(K,K).
\]
If $P=0$, the only supported directions are zero, and these expressions
are defined to be zero.
The square-root derivative used here is a case of the classical
divided-difference calculus for Hermitian matrix
functions~\cite[Section~1, Equation~(1)]{Daletskii1957}.
We give a direct inverse-function proof to keep the support conventions explicit.
\begin{lemma}[Energy derivatives]\label{lem:hessian}
On the cone of matrices positive definite on a fixed support,
$\Phi$ is smooth. For Hermitian directions $K,L$ on that support,
\[
d\Phi_P(K)=\frac32\tr(\sqrt P K),
\qquad
d^2\Phi_P(K,L)=\mathcal B_P(K,L).
\]
If $p_i>0$ are the eigenvalues on the support, then in an orthonormal
eigenbasis
\begin{equation}\label{eq:hessian-metric}
b_P(K)=\frac32\sum_{i,j}
\frac{|K_{ij}|^2}{\sqrt{p_i}+\sqrt{p_j}}.
\end{equation}
In particular, $\mathcal B_P$ is a positive definite symmetric bilinear
form on the real space of supported Hermitian directions, and
\[
|\mathcal B_P(K,L)|\le b_P(K)^{1/2}b_P(L)^{1/2}.
\]
\end{lemma}
\begin{proof}
Restrict to the support and put $A=\sqrt P$.
The derivative of the squaring map at $A$ is $X\mapsto AX+XA$.
In an eigenbasis of $A$, this map multiplies its $(i,j)$ entry by
$\sqrt{p_i}+\sqrt{p_j}>0$. The inverse function theorem therefore
gives a smooth positive square root. Its derivative $X_K$ in
direction $K$ satisfies
\[
AX_K+X_KA=K.
\]
Taking a trace after multiplication by $A$ yields
$\tr(PX_K)=\tfrac12\tr(AK)$. Differentiation of
$\Phi(P)=\tr(P\sqrt P)$ now gives
\[
d\Phi_P(K)=\tr(KA)+\tr(PX_K)=\frac32\tr(AK).
\]
Differentiating this identity in direction $L$ proves the Hessian
formula. The entrywise formula follows by diagonalizing $A$.
It also proves symmetry, positivity, and Cauchy--Schwarz on the
real space of Hermitian directions.
\end{proof}
The energy and its Hessian are unchanged by simultaneous unitary
conjugation of their arguments. They are also unchanged if all
arguments are extended by zero to a larger Hilbert space: the
positive eigenvalues and the supported entries in
\eqref{eq:hessian-metric} stay the same. We will use these
observations to compare Grams acting on different subspaces.
\Needspace{3\baselineskip}
\begin{lemma}[Order comparison]\label{lem:order}
Let $P,Q$ be positive semidefinite matrices on the same Hilbert space,
let $a\ge1$, and suppose $P\le aQ$. Every Hermitian $K$ supported
on $P$ is supported on $Q$, and
\begin{equation}\label{eq:matrix-order}
b_P(K)\ge a^{-1/2}b_Q(K).
\end{equation}
\end{lemma}
\begin{proof}
If $x\in\ker Q$, then $0\le\langle Px,x\rangle
\le a\langle Qx,x\rangle=0$, so $Px=0$.
Thus $\ker Q\subseteq\ker P$, proving the support assertion.
We recall the two operator-order facts needed for the comparison.
Inversion reverses order on positive definite matrices: for
$0<X\le Y$, conjugate by $X^{-1/2}$, invert the resulting matrix
at least $I$, and conjugate back. Moreover,
\[
\sqrt X=\frac1\pi\int_0^\infty
X(X+tI)^{-1}\,\frac{\dd t}{\sqrt t}.
\]
This norm-convergent formula follows from the scalar integral by
spectral calculus. Since
$X(X+tI)^{-1}=I-t(X+tI)^{-1}$, inversion order proves that the
positive square root preserves order. This is the classical
Loewner--Heinz square-root monotonicity; see, for example,
Kwong~\cite[Theorem~1 and Section~2]{Kwong1975}. The supported
Hessian comparison below is derived from these order facts here.
For $\eta>0$, the assumptions give
\[
P+\eta I\le a(Q+\eta I),
\qquad
\sqrt{P+\eta I}\le\sqrt a\,\sqrt{Q+\eta I}.
\]
If $A\le B$, then
$\mathsf L_A+\mathsf R_A\le\mathsf L_B+\mathsf R_B$ as
selfadjoint operators on Hilbert--Schmidt space. Indeed, left and
right multiplication by the positive matrix $B-A$ have the
nonnegative quadratic forms
$\tr(X^*(B-A)X)$ and $\tr(X^*X(B-A))$.
Inverting the resulting positive definite operators gives
\[
b_{P+\eta I}(K)\ge a^{-1/2}b_{Q+\eta I}(K).
\]
Finally let $\eta\downarrow0$.
Formula~\eqref{eq:hessian-metric}, applied in a full eigenbasis,
shows that both limits are the asserted supported costs: every
entry of $K$ meeting the kernel of the respective matrix is zero.
\end{proof}
We will apply the Hessian to Gram matrices with an insertion between
the two factors. The following bound controls such derivatives by
the energy itself.
\begin{lemma}[Inserted Gram bound]\label{lem:insertion}
Let $R:E\to F$ be a linear map between finite-dimensional complex
Hilbert spaces, and let $L:F\to F$ be Hermitian. The matrix $R^*LR$
is supported on $R^*R$, and
\begin{equation}\label{eq:insertion-bound}
b_{R^*R}(R^*LR)
\le\frac34\lVert L\rVert_{\op}^2\Phi(R^*R).
\end{equation}
\end{lemma}
\begin{proof}
The insertion annihilates $\ker R=\ker(R^*R)$ and is Hermitian,
so it is supported on the Gram. The assertion is immediate if
$R=0$. Otherwise take a singular-value decomposition, with positive
singular values $\rho_i$, and let $C$ be the compression of $L$ to
the corresponding left singular vectors. Then
$\lVert C\rVert_{\op}\le\lVert L\rVert_{\op}$ and
\[
b_{R^*R}(R^*LR)
=\frac32\sum_{i,j}
\frac{\rho_i^2\rho_j^2}{\rho_i+\rho_j}|C_{ij}|^2.
\]
For $x,y>0$, the arithmetic--geometric mean inequality gives
\[
\frac{x^2y^2}{x+y}
\le\frac{(xy)^{3/2}}2
\le\frac{x^3+y^3}4.
\]
Each row and each column of $C$ has squared Euclidean norm at most
$\lVert L\rVert_{\op}^2$. Consequently
\[
b_{R^*R}(R^*LR)
\le\frac38\sum_{i,j}(\rho_i^3+\rho_j^3)|C_{ij}|^2
\le\frac34\lVert L\rVert_{\op}^2\sum_i\rho_i^3,
\]
which is \eqref{eq:insertion-bound}.
\end{proof}
==== END The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/matrix.tex ====
==== BEGIN The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/maximal.tex ====
\section{From masked energies to the maximal transform}\label{sec:maximal}
The estimate of Section~\ref{sec:jumps} allows different truncation
endpoints at every output position. We first use this freedom on a finite
mesh, choosing the endpoints that maximize each output. We then pass to
functions and compare the resulting smooth kernels with hard truncations.
The comparison will be pointwise and uniform in both endpoints.
\subsection{Odd windows and selected endpoints}
For $s>0$ and $0<\varepsilon<R$, define
\[
k_s(t)=\frac{t}{\sqrt{s}}g_s(t),\qquad
\psi=k_1*k_1*k_1,\qquad
I_{\varepsilon,R}(t)=\int_{\varepsilon^2}^{R^2}
s^{-1/2}\psi(t/\sqrt{s})\,\frac{\dd s}{s}.
\]
Here $*$ is convolution on $\R$. The function $\psi$ is Schwartz, and
$I_{\varepsilon,R}$ is smooth and integrable for fixed positive endpoints.
\begin{proposition}[Selected smooth truncations]\label{prop:selected-kernel}
Let $h>0$, let $N\ge0$ be an integer, set $u_i=ih$, and let
$A_v(i,j)$, $v=0,1,2$, be complex
arrays indexed by $-N\le i,j\le N$. Assign an arbitrary endpoint pair
$h\le\varepsilon_{ij}<\rho_{ij}<\infty$ to each $(i,j)$. Then
\begin{equation}\label{eq:selected-kernel}
\left|h^3\sum_{i,j,l}A_0(i,j)A_1(j,l)A_2(l,i)
I_{\varepsilon_{ij},\rho_{ij}}(u_i+u_j+u_l)\right|
\le C\prod_{v=0}^2\left(\sum_{i,j}h^2|A_v(i,j)|^3\right)^{1/3}.
\end{equation}
The constant is independent of the arrays, mesh and selected endpoints.
\end{proposition}
\begin{proof}
We apply the mixed trace inequality with Gaussian weights of larger
variance than the windows $k_s$, so that the diagonal insertions are
bounded. We then compare each quadratic cost with a variance choice
covered by Proposition~\ref{prop:masked-dissipation}.
Use the masked array
$A_0^s(i,j)=A_0(i,j)\mathbf1_{\{\varepsilon_{ij}^2<s<\rho_{ij}^2\}}$
and the scale interval $[s_-,s_+]$ from
Proposition~\ref{prop:masked-dissipation}. All weighted matrices below use
this same mask. Let superscript $\mathrm b$ denote the base variances
$(\lambda s,\lambda s,\lambda s)$, where $\lambda=11/10$, and put
\[
D_v(i,i)=\frac{k_s(u_i-p_v)}{g_{\lambda s}(u_i-p_v)}.
\]
These matrices are real diagonal. With $z=(u_i-p_v)/\sqrt{s}$,
\[
|D_v(i,i)|=\sqrt\lambda\,|z|
e^{-(\lambda-1)z^2/(2\lambda)}
\le\frac{\lambda}{\sqrt{e(\lambda-1)}}.
\]
Thus their operator norms have an absolute bound.
For each vertex $v$, let $\boldsymbol\alpha(v)$ equal $1$ at $v$ and
$\lambda$ at the other two vertices. In the star centered at $v$, changing
from the base variances to $\boldsymbol\alpha(v)s$ changes only its row
weight. In particular,
\[
R_v^{\boldsymbol\alpha(v)}
=\diag_i\left(\sqrt{\frac{g_s(u_i-p_v)}
{g_{\lambda s}(u_i-p_v)}}\right)R_v^{\mathrm b},
\qquad
T_v^{\boldsymbol\alpha(v)}\le\sqrt\lambda\,T_v^{\mathrm b}.
\]
The last inequality uses $g_s\le\sqrt\lambda\,g_{\lambda s}$.
The score $L_v^{\boldsymbol\alpha(v)}$ equals
$\diag_i((u_i-p_v)/s)$, so the row densities also give the exact identity
\begin{equation}\label{eq:maximal-window-score}
(R_v^{\mathrm b})^*D_vR_v^{\mathrm b}
=\sqrt{s}\,(R_v^{\boldsymbol\alpha(v)})^*
L_v^{\boldsymbol\alpha(v)}R_v^{\boldsymbol\alpha(v)}.
\end{equation}
Indeed, both inserted row densities are $h k_s(u_i-p_v)$, and all column
weights agree. The insertion on the left is therefore supported on
$T_v^{\boldsymbol\alpha(v)}$. Lemma~\ref{lem:order} and quadratic
homogeneity give
\[
b_{T_v^{\mathrm b}}\bigl((R_v^{\mathrm b})^*D_vR_v^{\mathrm b}\bigr)
\le\lambda^{1/4}s V_v^{\boldsymbol\alpha(v)}.
\]
Apply Lemma~\ref{lem:mixed-trace} to the base matrices. Integrate its three
nonnegative terms separately, using Proposition~\ref{prop:masked-dissipation}
with the corresponding $\boldsymbol\alpha(v)$. The factor $s$ cancels
$\dd s/s$, and we obtain
\begin{equation}\label{eq:masked-window-integral}
\int_{s_-}^{s_+}\int_{\Pi_0}
\left|\tr(D_0W_0^{\mathrm b}D_1W_1^{\mathrm b}D_2W_2^{\mathrm b})\right|
\,\dd\pi_0\,\frac{\dd s}{s}
\le C\mathcal S(A).
\end{equation}
We identify the signed integral in this estimate. Expanding the trace
leaves one signed Gaussian weight at each vertex:
\[
h^3\sum_{i,j,l}A_0^s(i,j)A_1(j,l)A_2(l,i)
k_s(u_i-p_0)k_s(u_j-p_1)k_s(u_l-p_2).
\]
On $\Pi_0$, substitute $r=u_i-p_0$ and $q=u_j-p_1$.
The Jacobian is one and the third argument becomes
$u_i+u_j+u_l-r-q$. Consequently the center integral is
$(k_s*k_s*k_s)(u_i+u_j+u_l)$. Since
$k_s(t)=s^{-1/2}k_1(t/\sqrt{s})$, this convolution equals
\[
s^{-1/2}\psi\bigl((u_i+u_j+u_l)/\sqrt{s}\bigr).
\]
The sums are finite and the Gaussian products are absolutely integrable
on the center plane and the compact positive scale interval. Thus all
these integrations can be performed term by term. The mask restricts the
scale integral of the $(i,j,l)$ term to
$(\varepsilon_{ij}^2,\rho_{ij}^2)$. Taking the absolute value after
integration in \eqref{eq:masked-window-integral} proves
\eqref{eq:selected-kernel} with its right side initially replaced by
$C\mathcal S(A)$.
Put $n_v=(\sum h^2|A_v|^3)^{1/3}$. If any $n_v$ is zero the signed sum
vanishes. Otherwise apply the last estimate to $A_v/n_v$ for all three
vertices, keeping the selected endpoints fixed. The normalized value of
$\mathcal S$ is $3$. Scaling the signed trilinear expression back proves
the proposition.
\end{proof}
\subsection{Finite maxima and the continuum limit}
For functions $F,G$ on $\R^2$, write
\[
B^I_{\varepsilon,R}(F,G)(x,y)
=\int_\R F(x+t,y)G(x,y+t)I_{\varepsilon,R}(t)\,\dd t.
\]
For compactly supported smooth inputs this integral is defined at every
point. We now place the supremum over its endpoints inside the spatial
norm.
\begin{proposition}\label{prop:smooth-kernel-maximal}
For all complex $F,G\in C_c^\infty(\R^2)$,
\[
\left\|\sup_{0<\varepsilon<R}
|B^I_{\varepsilon,R}(F,G)|\right\|_{L^{3/2}(\R^2)}
\le C\|F\|_{L^3(\R^2)}\|G\|_{L^3(\R^2)}.
\]
The supremum is measurable and the constant is absolute.
\end{proposition}
\begin{proof}
First fix a nonempty finite family of endpoint pairs
$(\varepsilon_k,R_k)$, $1\le k\le m$, and take
$h\le\min_k\varepsilon_k$. For finite arrays $A_1,A_2$, define
\[
U_k(i,j)=h\sum_l A_1(j,l)A_2(l,i)
I_{\varepsilon_k,R_k}(u_i+u_j+u_l),
\qquad M(i,j)=\max_{1\le k\le m}|U_k(i,j)|.
\]
At each output pair choose the least index attaining this maximum, and
write $U(i,j)$ for its output. These choices provide admissible
$\varepsilon_{ij},\rho_{ij}$ in Proposition~\ref{prop:selected-kernel}.
They are fixed while the testing array $A_0$ varies.
Set $H=\sum_{i,j}h^2 M(i,j)^{3/2}$. If $H=0$ the desired bound is
immediate. Otherwise choose
\[
A_0(i,j)=
\begin{cases}
H^{-1/3}M(i,j)^{1/2}\,\overline{U(i,j)}/|U(i,j)|,
&U(i,j)\ne0,\\
0,&U(i,j)=0.
\end{cases}
\]
Then $\sum h^2|A_0|^3=1$ and $h^2\sum A_0U=H^{2/3}$.
Proposition~\ref{prop:selected-kernel} therefore gives
\begin{equation}\label{eq:discrete-finite-maximum}
\left(\sum_{i,j}h^2M(i,j)^{3/2}\right)^{2/3}
\le C\prod_{v=1}^2\left(\sum_{i,j}h^2|A_v(i,j)|^3\right)^{1/3}.
\end{equation}
This estimate has no dependence on the number of endpoint pairs. The
complex phase of $A_0$ accounts for the absolute value at each output.
To pass to functions, define the compactly supported smooth inputs
\[
G_1(b,c)=F(b+c,-b),\qquad G_2(c,a)=G(-a,c+a).
\]
Both pair maps have determinant one, so
$\|G_1\|_3=\|F\|_3$ and $\|G_2\|_3=\|G\|_3$.
Choose $B>0$ so that both supports lie in the interior of $[-B,B]^2$.
Take $h=B/N$, $-N\le i,j,l\le N$, and sample
\[
A_1(j,l)=G_1(jh,lh),\qquad A_2(l,i)=G_2(lh,ih).
\]
The corresponding continuum outputs are
\[
\mathcal U_k(a,b)=\int_\R G_1(b,c)G_2(c,a)
I_{\varepsilon_k,R_k}(a+b+c)\,\dd c.
\]
They vanish outside $[-B,B]^2$: $G_2(c,a)$ restricts $a$ and $G_1(b,c)$
restricts $b$. A nonzero integrand also has $c\in[-B,B]$.
For our fixed finite endpoint family the integrands and their $c$
derivatives are uniformly bounded on this compact cube. Riemann sums
in $c$ consequently converge uniformly for $(a,b)\in[-B,B]^2$ and
$1\le k\le m$. In particular,
\[
\max_{i,j}\max_{1\le k\le m}
|U_k(i,j)-\mathcal U_k(ih,jh)|\longrightarrow0.
\]
The finite maximum of magnitudes preserves this uniform convergence.
Since $\max_k|\mathcal U_k|$ is continuous and compactly supported,
the two-dimensional Riemann sums on the left of
\eqref{eq:discrete-finite-maximum} converge to its $L^{3/2}$ norm.
The cube roots of the input sums converge to $\|G_1\|_3$ and
$\|G_2\|_3$.
The condition on $h$ holds for all sufficiently large $N$.
We obtain
\[
\|\max_k|\mathcal U_k|\|_{3/2}\le C\|G_1\|_3\|G_2\|_3.
\]
No convergence or regularity of the maximizing indices is needed: we
passed to the limit in the finite maximum itself.
With $x=-a$, $y=-b$, and $t=a+b+c$, the last integral is
$B^I_{\varepsilon_k,R_k}(F,G)(x,y)$. The output coordinate change
preserves Lebesgue measure, giving the asserted estimate for every
finite endpoint family.
Finally enumerate all rational pairs $0<\varepsilon<R$ and take
increasing finite subfamilies. At each spatial point the output is
continuous in the two positive endpoints, by its defining integrals
on compact scale and spatial intervals. Thus the supremum over these
rational pairs equals the supremum over all real pairs. It is measurable
as a countable supremum of continuous spatial functions. Monotone
convergence of the $3/2$ powers of the finite maxima proves the result.
\end{proof}
\subsection{A pointwise comparison with hard truncations}
The smooth maximal estimate will imply the hard one because each
endpoint error has a common pointwise majorant. A uniform bound on
its kernel's $L^1$ norm alone would not provide this comparison.
\begin{lemma}\label{lem:maximal-hard-kernel}
There are an absolute constant $C$ and functions $E_r$, $r>0$, such that,
for $t\ne0$, apart from endpoint values,
\begin{equation}\label{eq:maximal-hard-decomposition}
I_{\varepsilon,R}(t)
=\frac{c_0}{t}\mathbf1_{\{\varepsilon<|t|<R\}}
+E_\varepsilon(t)-E_R(t),
\qquad c_0=-\frac{2}{3\sqrt{6\pi}},
\end{equation}
and
\begin{equation}\label{eq:maximal-error-envelope}
|E_r(t)|\le\frac{C}{r}\left(1+\frac{|t|}{r}\right)^{-2}
\qquad(t\ne0).
\end{equation}
\end{lemma}
\begin{proof}
Since $k_1=-g_1'$ and Gaussian variances add under convolution,
\[
\psi(t)=-g_3'''(t)
=\left(\frac{t^3}{27}-\frac t3\right)g_3(t).
\]
The convolution identity follows by completing the square;
differentiation under convolution is justified by Schwartz decay.
For $v\ge0$, put
\[
P(v)=2\int_0^v\psi(z)\,\dd z
=2g_3''(0)-2g_3''(v),\qquad
c_0=P(\infty)=2g_3''(0)=-\frac2{3\sqrt{6\pi}}.
\]
The substitution $z=t/\sqrt{s}$ for $t>0$, followed by oddness for
$t<0$, yields
\[
I_{\varepsilon,R}(t)
=\frac{P(|t|/\varepsilon)-P(|t|/R)}{t}.
\]
Define $Q(v)=P(v)-c_0\mathbf1_{\{v>1\}}$ for $v\ge0$.
Smooth oddness of $\psi$ gives $P(v)=O(v^2)$ at zero. For $v>1$,
\[
Q(v)=-2g_3''(v)
=-2\left(\frac{v^2}{9}-\frac13\right)g_3(v).
\]
Thus $Q$ has Gaussian decay at infinity and bounded one-sided values
at its sole jump, $v=1$. Enlarging an absolute constant on a compact
interval, we have
\[
\frac{|Q(v)|}{v}\le C(1+v)^{-2}\qquad(v>0).
\]
Set $E_r(t)=Q(|t|/r)/t$ for $t\ne0$ and set $E_r(0)=0$.
Substituting the definition of $Q$ proves
\eqref{eq:maximal-hard-decomposition} and
\eqref{eq:maximal-error-envelope}. Values at the four hard endpoints
have no effect on any integral.
\end{proof}
Let $M_x$ and $M_y$ denote the continuous centered interval maximal
operators in the first and second variables, respectively. The next
estimate follows from their one-dimensional bounds in
Lemma~\ref{lem:line-maximal}.
\begin{lemma}\label{lem:pointwise-endpoint-error}
For complex $F,G\in C_c^\infty(\R^2)$ and every $r>0$,
\[
\left|\int_\R F(x+t,y)G(x,y+t)E_r(t)\,\dd t\right|
\le C\bigl[M_x(|F|^2)(x,y)M_y(|G|^2)(x,y)\bigr]^{1/2}.
\]
The right side has $L^{3/2}(\R^2)$ norm at most $C\|F\|_3\|G\|_3$.
\end{lemma}
\begin{proof}
Put $q_r(t)=r^{-1}(1+|t|/r)^{-2}$. For nonnegative locally integrable
$f$ on $\R$, the part of $\int q_r(t)f(x+t)\,\dd t$ with $|t|\le r$
is at most $2M_{\R}f(x)$. On the shell
$2^kr<|t|\le2^{k+1}r$, $k\ge0$, its contribution is at most
\[
r^{-1}2^{-2k}\int_{|t|\le2^{k+1}r}f(x+t)\,\dd t
\le 4\,2^{-k}M_{\R}f(x).
\]
Summing the geometric series proves a bound independent of $r$.
By \eqref{eq:maximal-error-envelope} and weighted Cauchy--Schwarz,
\begin{align*}
\left|\int F(x+t,y)G(x,y+t)E_r(t)\,\dd t\right|
&\le C\left(\int |F(x+t,y)|^2q_r(t)\,\dd t\right)^{1/2}\\
&\hspace{12mm}\cdot
\left(\int |G(x,y+t)|^2q_r(t)\,\dd t\right)^{1/2}.
\end{align*}
Apply the preceding one-dimensional bound on the respective coordinate
lines. This gives the pointwise assertion.
H\"older's inequality and Lemma~\ref{lem:line-maximal}, applied on slices
and then integrated in the other coordinate, give
\begin{align*}
\|[M_x(|F|^2)M_y(|G|^2)]^{1/2}\|_{3/2}
&\le\|M_x(|F|^2)\|_{3/2}^{1/2}
\|M_y(|G|^2)\|_{3/2}^{1/2}\\
&\le C\||F|^2\|_{3/2}^{1/2}\||G|^2\|_{3/2}^{1/2}
=C\|F\|_3\|G\|_3.
\end{align*}
\end{proof}
\begin{proposition}[Compact smooth inputs]\label{prop:compact-maximal}
For all complex $F,G\in C_c^\infty(\R^2)$,
\[
\|B_*(F,G)\|_{L^{3/2}(\R^2)}
\le C\|F\|_{L^3(\R^2)}\|G\|_{L^3(\R^2)},
\]
with an absolute constant. The supremum defining $B_*$ is over both
finite hard truncation endpoints, inside the spatial norm.
\end{proposition}
\begin{proof}
Lemma~\ref{lem:maximal-hard-kernel} and
Lemma~\ref{lem:pointwise-endpoint-error} give, at each spatial point,
\[
|c_0|\sup_{0<\varepsilon<R}|B_{\varepsilon,R}(F,G)|
\le \sup_{0<\varepsilon<R}|B^I_{\varepsilon,R}(F,G)|
+2C[M_x(|F|^2)M_y(|G|^2)]^{1/2}.
\]
Both endpoint errors have the same majorant, independently of their
radii. The hard outputs are continuous in both positive endpoints,
so their supremum also equals the measurable supremum over rational
pairs. Take $L^{3/2}$ norms, apply
Proposition~\ref{prop:smooth-kernel-maximal} and
Lemma~\ref{lem:pointwise-endpoint-error}, and divide by $|c_0|>0$.
\end{proof}
==== END The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/maximal.tex ====
==== BEGIN The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/trace.tex ====
\section{A mixed trace inequality}\label{sec:trace}
The triangular form is naturally represented by a cyclic product of
operators. Such a representation also appears in the dyadic model of
Kova\v{c}, Thiele and Zorin-Kranich~\cite[Equation~(1.5)]{KovacThieleZorinKranich2015}.
We prove a finite-dimensional estimate that controls this product by
three of the quadratic costs from Section~\ref{sec:matrix}.
Let $H_v$, $v\in\mathbb Z/3\mathbb Z$, be finite-dimensional complex
Hilbert spaces, and let $W_v:H_{v+1}\to H_v$ be arbitrary linear maps.
Combine the two edges incident to $v$ into the \emph{row star}
\begin{equation}\label{eq:row-star}
R_v=[\,W_v\quad W_{v-1}^*\,]:
H_{v+1}\oplus H_{v-1}\longrightarrow H_v.
\end{equation}
Its two Gram matrices are
\begin{equation}\label{eq:star-grams}
T_v=R_v^*R_v,\qquad S_v=R_vR_v^*.
\end{equation}
Their positive eigenvalues are the squares of the nonzero singular
values of $R_v$, with the same multiplicities. In particular,
$\Phi(T_v)=\Phi(S_v)$.
Figure~\ref{fig:triangle} shows the cyclic edges and the row star at
vertex $0$.
\input{figures/triangle}
\begin{lemma}[Mixed trace estimate]\label{lem:mixed-trace}
Let $W_v:H_{v+1}\to H_v$ be linear maps between finite-dimensional
complex Hilbert spaces, with row stars $R_v$ and Gram matrices $T_v$
as in \eqref{eq:row-star}--\eqref{eq:star-grams}.
For Hermitian maps $D_v:H_v\to H_v$, put
$D_* =\max_v\lVert D_v\rVert_{\op}$. Then
\begin{equation}\label{eq:mixed-trace}
\left|\tr(D_0W_0D_1W_1D_2W_2)\right|
\le \frac53D_*\sum_{v=0}^2
b_{T_v}(R_v^*D_vR_v).
\end{equation}
The constant is independent of the dimensions of the spaces.
\end{lemma}
\begin{proof}
We first put the three quadratic costs in a common metric. We then
estimate a cubic trace in an eigenbasis and recover the absolute value
of the original cyclic product by changing the phase of one edge.
\paragraph{A common quadratic form.}
On $H=H_0\oplus H_1\oplus H_2$, define the Hermitian matrices
\[
M=\begin{pmatrix}
0&W_0&W_2^*\\
W_0^*&0&W_1\\
W_2&W_1^*&0
\end{pmatrix},
\qquad D=\diag(D_0,D_1,D_2).
\]
Thus $\lVert D\rVert_{\op}=D_*$. Let $P_v$ be the orthogonal
projection onto $H_v$ and set
\[
A_v=MP_vM,\qquad K_v=MP_vDM.
\]
Only the $v$-th block row of $P_vM$ can be nonzero; it consists of
$R_v$ with its columns placed in the corresponding vertex spaces.
Consequently $A_v$ and $K_v$ are,
respectively, $T_v$ and $R_v^*D_vR_v$, after a permutation of blocks
and extension by zero. These operations preserve the quadratic cost.
Since $P_v$ commutes with $D$, the matrix $K_v$ is Hermitian. Moreover,
\[
\ker A_v=\ker(P_vM)\subseteq\ker K_v,
\qquad 0\le A_v\le M^2.
\]
The first identity follows from
$\langle A_vx,x\rangle=\lVert P_vMx\rVert^2$.
Thus $K_v$ is supported on $A_v$, and hence also on $M^2$.
Lemma~\ref{lem:order}, followed by the quadratic-form inequality for
three vectors, yields
\begin{equation}\label{eq:trace-common-metric}
\begin{aligned}
\mathcal E
&:=\sum_v b_{T_v}(R_v^*D_vR_v)\\
&\ge\sum_v b_{M^2}(K_v)
\ge\frac13 b_{M^2}\left(\sum_vK_v\right)
=\frac13 b_{M^2}(MDM).
\end{aligned}
\end{equation}
All these quadratic forms are taken on their supports; no
invertibility of $M$ is required.
\paragraph{The cubic trace.}
Choose an orthonormal eigenbasis $(e_i)$ of $M$, write the real
eigenvalues as $m_i$, and arrange the indices so that $q_i:=|m_i|$ is
nonincreasing. Equal values may be ordered arbitrarily. Write $D_{ij}$
for the entries of $D$ in this basis. The spectral formula for
$b_{M^2}$ gives
\begin{equation}\label{eq:trace-Q}
\mathcal E\ge\frac12Q,
\qquad
Q:=\sum_{i,j:q_i+q_j>0}
\frac{q_i^2q_j^2}{q_i+q_j}|D_{ij}|^2.
\end{equation}
If either eigenvalue is zero, its summand is zero; pairs of zero
eigenvalues are omitted. Define
\[
A_i=q_i^3|D_{ii}|^2,\qquad
E_i=q_i\sum_{j>i}q_j^2|D_{ij}|^2,
\qquad A=\sum_iA_i,\quad E=\sum_iE_i.
\]
Hermitian symmetry and the ordering of the $q_i$ imply
\begin{equation}\label{eq:trace-AE}
\begin{aligned}
Q
&=\frac A2+
2\sum_{\substack{i<j\\q_i+q_j>0}}
\frac{q_i^2q_j^2}{q_i+q_j}|D_{ij}|^2
\ge\frac A2+E.
\end{aligned}
\end{equation}
Indeed, for $i<j$ we have $q_j\le q_i$, so
$2q_i/(q_i+q_j)\ge1$ whenever the denominator is positive.
We now bound
\[
\tr((DM)^3)
=\sum_{i,j,k}m_im_jm_kD_{ij}D_{jk}D_{ki}
\]
by grouping the triples according to how often their smallest index
occurs. This is the index of a largest absolute eigenvalue in the
triple.
If the smallest index $i$ occurs once, its three cyclic positions
give
\[
3m_i\sum_{j,k>i}m_jm_kD_{ij}D_{jk}D_{ki}.
\]
Here $j=k$ is allowed. For the vector $x=(m_jD_{ji})_{j>i}$, the inner
sum is $x^*D_{>i}x$, where $D_{>i}$ is the compression of $D$ to the
indices greater than $i$. Since
$\lVert D_{>i}\rVert_{\op}\le D_*$, the modulus of this contribution
is at most $3D_*E_i$. This argument retains the signs of the $m_j$
and uses $D_{ji}=\overline{D_{ij}}$.
If the smallest index $i$ occurs twice, the contribution is
\[
3m_i^2D_{ii}\sum_{k>i}m_k|D_{ik}|^2.
\]
Cauchy--Schwarz and
$\sum_k|D_{ik}|^2=\lVert De_i\rVert^2\le D_*^2$ give
\[
\left|\sum_{k>i}m_k|D_{ik}|^2\right|
\le D_*\left(\sum_{k>i}q_k^2|D_{ik}|^2\right)^{1/2}.
\]
Its contribution therefore has modulus at most
$3D_*\sqrt{A_iE_i}$. This also holds when $q_i=0$, without any
division by $q_i$.
Finally, when all three indices equal $i$, the modulus is at most
$D_*A_i$. These three cases exhaust all triples, including repeated
indices. Summing them and applying Cauchy--Schwarz once more, we obtain
\begin{equation}\label{eq:trace-cubic}
\begin{aligned}
|\tr((DM)^3)|
&\le D_*\bigl(A+3E+3\sqrt{AE}\bigr)\\
&\le D_*\left(\frac52A+\frac92E\right)
\le 5D_*\left(\frac A2+E\right)
\le5D_*Q.
\end{aligned}
\end{equation}
\paragraph{Recovering the complex cyclic product.}
We have bounded the cubic trace by the star costs. It remains to
extract the modulus of the desired triangular trace. Put
\[
\tau=\tr(D_0W_0D_1W_1D_2W_2).
\]
Because $M$ has zero diagonal blocks, a nonzero closed three-step
block product must visit all three vertices. The three starting
points in one orientation each give $\tau$, by cyclicity of the
trace. Reversing the orientation gives $\overline\tau$, since the
$D_v$ are Hermitian. Hence
\[
\tr((DM)^3)=6\operatorname{Re}\tau.
\]
If $\tau\ne0$, choose $z\in\mathbb C$ with $|z|=1$ and
$z\tau=|\tau|$, and replace $W_0$ by $zW_0$. The row stars change by
\[
\begin{aligned}
R_0&\longmapsto
R_0\diag(zI_{H_1},I_{H_2}),\\
R_1&\longmapsto
R_1\diag(I_{H_2},\overline zI_{H_0}),
\end{aligned}
\]
while $R_2$ is unchanged. Each displayed right multiplier is unitary
and conjugates the corresponding Gram and inserted Gram by the same
unitary. Thus $\mathcal E$ is unchanged, as is $D_*$.
Applying \eqref{eq:trace-Q} and \eqref{eq:trace-cubic} to the modified
matrices, with $Q$ computed for the new $M$, gives
\[
6|\tau|\le5D_*Q\le10D_*\mathcal E,
\]
which is \eqref{eq:mixed-trace}. The case $\tau=0$ is immediate.
\end{proof}
==== END The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//sections/trace.tex ====
==== BEGIN The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//references.bib ====
@misc{GrafakosEtAl2017,
author = {Grafakos, Loukas and Oliveira e Silva, Diogo and Pramanik, Malabika and Seeger, Andreas and Stovall, Betsy},
title = {Some problems in harmonic analysis},
year = {2017},
howpublished = {\href{https://arxiv.org/abs/1701.06637v1}{arXiv:1701.06637v1}},
eprint = {1701.06637},
archivePrefix = {arXiv},
primaryClass = {math.CA},
url = {https://arxiv.org/abs/1701.06637v1},
note = {Section 9, contributed by Christoph Thiele, Problem 13}
}
@article{KovacThieleZorinKranich2015,
author = {Kova\v{c}, Vjekoslav and Thiele, Christoph and Zorin-Kranich, Pavel},
title = {Dyadic triangular {Hilbert} transform of two general functions and one not too general function},
journal = {Forum of Mathematics, Sigma},
volume = {3},
pages = {e25},
year = {2015},
doi = {10.1017/fms.2015.25},
url = {https://arxiv.org/abs/1506.00861v1},
note = {Version consulted: arXiv:1506.00861v1}
}
@article{ZorinKranich2017,
author = {Zorin-Kranich, Pavel},
title = {Cancellation for the simplex {Hilbert} transform},
journal = {Mathematical Research Letters},
volume = {24},
number = {2},
pages = {581--592},
year = {2017},
doi = {10.4310/MRL.2017.v24.n2.a16},
url = {https://arxiv.org/abs/1507.02436v1},
note = {Version consulted: arXiv:1507.02436v1}
}
@article{DurcikKovacSkrebThiele2019,
author = {Durcik, Polona and Kova\v{c}, Vjekoslav and {\v S}kreb, Kristina Ana and Thiele, Christoph},
title = {Norm-variation of ergodic averages with respect to two commuting transformations},
journal = {Ergodic Theory and Dynamical Systems},
volume = {39},
number = {3},
pages = {658--688},
year = {2019},
doi = {10.1017/etds.2017.48},
url = {https://arxiv.org/abs/1603.00631v3},
note = {Version consulted: \href{https://arxiv.org/abs/1603.00631v3}{arXiv:1603.00631v3}}
}
@article{DurcikKovacThiele2019,
author = {Durcik, Polona and Kova\v{c}, Vjekoslav and Thiele, Christoph},
title = {Power-type cancellation for the simplex {Hilbert} transform},
journal = {Journal d'Analyse Math\'ematique},
volume = {139},
pages = {67--82},
year = {2019},
doi = {10.1007/s11854-019-0052-4},
url = {https://arxiv.org/abs/1608.00156v2},
note = {Version consulted: \href{https://arxiv.org/abs/1608.00156v2}{arXiv:1608.00156v2}}
}
@article{DurcikRoos2021,
author = {Durcik, Polona and Roos, Joris},
title = {Averages of simplex {Hilbert} transforms},
journal = {Proceedings of the American Mathematical Society},
volume = {149},
number = {2},
pages = {633--647},
year = {2021},
doi = {10.1090/proc/15196},
url = {https://arxiv.org/abs/1812.11701v3},
note = {Version consulted: \href{https://arxiv.org/abs/1812.11701v3}{arXiv:1812.11701v3}}
}
@article{Kovac2011,
author = {Kova\v{c}, Vjekoslav},
title = {{Bellman} function technique for multilinear estimates and an application to generalized paraproducts},
journal = {Indiana University Mathematics Journal},
volume = {60},
number = {3},
pages = {813--846},
year = {2011},
doi = {10.1512/iumj.2011.60.4784},
url = {https://arxiv.org/abs/1108.0917v1},
note = {Version consulted: \href{https://arxiv.org/abs/1108.0917v1}{arXiv:1108.0917v1}}
}
@article{Durcik2015,
author = {Durcik, Polona},
title = {An {$L^4$} estimate for a singular entangled quadrilinear form},
journal = {Mathematical Research Letters},
volume = {22},
number = {5},
pages = {1317--1332},
year = {2015},
url = {https://arxiv.org/abs/1412.2384v2},
note = {Version consulted: \href{https://arxiv.org/abs/1412.2384v2}{arXiv:1412.2384v2}}
}
@misc{LinSlavikova2026,
author = {Lin, Fred Yu-Hsiang and Slav\'ikov\'a, Lenka},
title = {Rough averages of triangular {Hilbert} transforms},
year = {2026},
eprint = {2607.20206},
archivePrefix = {arXiv},
primaryClass = {math.CA},
url = {https://arxiv.org/abs/2607.20206v1},
note = {\href{https://arxiv.org/abs/2607.20206v1}{arXiv:2607.20206v1}, 22 July 2026}
}
@misc{HsuLin2026,
author = {Hsu, Martin and Lin, Fred Yu-Hsiang},
title = {Smoothing inequalities for corner-type bilinear averages: geometric characterization and applications},
year = {2026},
eprint = {2410.15791},
archivePrefix = {arXiv},
primaryClass = {math.CA},
url = {https://arxiv.org/abs/2410.15791v2},
note = {\href{https://arxiv.org/abs/2410.15791v2}{arXiv:2410.15791v2}, 18 July 2026}
}
@article{ChristDurcikRoos2021,
author = {Christ, Michael and Durcik, Polona and Roos, Joris},
title = {Trilinear smoothing inequalities and a variant of the triangular {Hilbert} transform},
journal = {Advances in Mathematics},
volume = {390},
pages = {107863},
year = {2021},
doi = {10.1016/j.aim.2021.107863},
url = {https://arxiv.org/abs/2008.10140v2},
note = {Version consulted: arXiv:2008.10140v2}
}
@article{Kovac2012Twisted,
author = {Kova\v{c}, Vjekoslav},
title = {Boundedness of the twisted paraproduct},
journal = {Revista Matem\'atica Iberoamericana},
volume = {28}, number = {4}, pages = {1143--1164}, year = {2012},
doi = {10.4171/RMI/707},
url = {https://arxiv.org/abs/1011.6140v1},
note = {Version consulted: arXiv:1011.6140v1; Lemma 2}
}
@article{Lewis1995,
author = {Lewis, Adrian S.},
title = {The convex analysis of unitarily invariant matrix functions},
journal = {Journal of Convex Analysis},
volume = {2}, number = {1--2}, pages = {173--183}, year = {1995},
url = {https://people.orie.cornell.edu/aslewis/publications/95-convex.pdf}
}
@article{HardyLittlewood1930,
author = {Hardy, G. H. and Littlewood, J. E.},
title = {A maximal theorem with function-theoretic applications},
journal = {Acta Mathematica},
volume = {54}, pages = {81--116}, year = {1930},
url = {https://archive.ymsc.tsinghua.edu.cn/pacm_download/117/5446-11511_2006_Article_BF02547518.pdf}
}
@article{DemeterThiele2010,
author = {Demeter, Ciprian and Thiele, Christoph},
title = {On the two-dimensional bilinear {Hilbert} transform},
journal = {American Journal of Mathematics},
volume = {132},
number = {1},
pages = {201--256},
year = {2010},
doi = {10.1353/ajm.0.0101},
eprint = {0803.1268},
archivePrefix = {arXiv},
primaryClass = {math.CA},
url = {https://arxiv.org/abs/0803.1268v1},
note = {Version consulted: arXiv:0803.1268v1, Section 6}
}
@article{Tao2016Cancellation,
author = {Tao, Terence},
title = {Cancellation for the multilinear {Hilbert} transform},
journal = {Collectanea Mathematica},
volume = {67},
number = {2},
pages = {191--206},
year = {2016},
eprint = {1505.06479},
archivePrefix = {arXiv},
primaryClass = {math.CA},
url = {https://arxiv.org/abs/1505.06479v3},
note = {Version consulted: arXiv:1505.06479v3}
}
@article{Daletskii1957,
author = {Daletskii, Yu. L.},
title = {Integration and differentiation of functions of {Hermitian} operators depending on a parameter},
journal = {Uspekhi Matematicheskikh Nauk},
volume = {12}, number = {1(73)}, pages = {182--186}, year = {1957},
url = {https://www.mathnet.ru/eng/rm7543},
note = {In Russian; Section 1, formula (1)}
}
@article{Kwong1975,
author = {Kwong, Man Kam},
title = {Inequalities for the powers of nonnegative {Hermitian} operators},
journal = {Proceedings of the American Mathematical Society},
volume = {51},
number = {2},
pages = {401--406},
year = {1975},
doi = {10.1090/S0002-9939-1975-0374970-X},
url = {https://doi.org/10.1090/S0002-9939-1975-0374970-X}
}
==== END The-maximal-triangular-Hilbert-transform-at-the-symmetric-point//references.bib ====