\documentclass[11pt]{article} \usepackage{microtype,amsmath,amssymb,amsthm,mathtools,booktabs,array} \usepackage[margin=1.05in]{geometry} \usepackage[colorlinks=true,linkcolor=blue!45!black,citecolor=blue!45!black,urlcolor=blue!45!black]{hyperref} \usepackage{xcolor} \hypersetup{pdftitle={The dimension-degree boundary for Gaussian moment vanishing},pdfsubject={Gaussian moment vanishing and the dimension-degree boundary},bookmarksdepth=2} \newtheorem{theorem}{Theorem}[section] \newtheorem{proposition}[theorem]{Proposition} \newtheorem{lemma}[theorem]{Lemma} \newtheorem{corollary}[theorem]{Corollary} \theoremstyle{definition}\newtheorem{definition}[theorem]{Definition} \theoremstyle{remark}\newtheorem{remark}[theorem]{Remark} \newcommand{\E}{\mathbb E} \newcommand{\C}{\mathbb C} \newcommand{\tr}{\operatorname{tr}} \newcommand{\wt}{\operatorname{wt}} \newcommand{\G}{\mathsf G} \DeclareMathOperator{\Span}{span} \setcounter{tocdepth}{1} \title{The dimension--degree boundary\\for Gaussian moment vanishing} \author{Roy van Rijn\\{\small Independent researcher}} \date{17 November 2026} \begin{document} \maketitle \begin{abstract} We classify complex polynomials of degree at most three in three independent standard real Gaussian variables whose positive moments all vanish. Up to a complex orthogonal linear change of coordinates and a nonzero scalar, every such polynomial has one-sided circular support or belongs to a single exceptional orbit. For every fixed polynomial $Q$, its mixed moments $\E[QP^m]$ vanish whenever $m>3\deg Q$, and the coefficient three is optimal. The proof extracts a Gaussian coordinate in which the polynomial is affine linear, then classifies that family using formal stationary frames and analytic period estimates. Combining this ternary-cubic result with Wilson's two-variable theorem, Liu's quadratic theorem and Long's counterexamples gives the exact dimension--degree boundary for the Gaussian Moments Conjecture: the implication from pure-moment vanishing to eventual mixed-moment vanishing holds universally precisely when there are at most two variables, the degree is at most two, or there are three variables and the degree is at most three. We reproduce the all-order moment identities for the two minimal failures. \end{abstract} \clearpage \tableofcontents \clearpage \section{What vanishing moments tell us}\label{sec:introduction} Let $X_1,\ldots,X_n$ be independent standard real Gaussian random variables. We allow a polynomial $P(X)$ to have complex coefficients. Suppose that \begin{equation}\label{eq:pure} \E[P(X)^m]=0\qquad\text{for every integer }m\geq1. \end{equation} The Gaussian Moments Conjecture asks whether this forces \begin{equation}\label{eq:safe} \E[Q(X)P(X)^m]=0\qquad\text{for all sufficiently large }m \end{equation} for every fixed polynomial $Q$. The threshold may depend on both $P$ and $Q$. The complex coefficients matter. For real $P$, the second moment in \eqref{eq:pure} forces $P=0$. For complex $P$, cancellation can make every positive moment vanish even when $P$ is far from zero. For example, $(X_1+iX_2)/\sqrt2$ has this property. The question is whether a fixed polynomial multiplier can keep exposing a nonzero moment at arbitrarily large powers. We call \eqref{eq:pure} \emph{purity} and \eqref{eq:safe}, for every $Q$, \emph{safety}. These names distinguish the hypothesis from the conclusion; safety without purity is a separate question. In the original formulation of Derksen, van den Essen and Zhao~\cite{DVEZ}, the assertion that every pure polynomial is safe was a conjecture in every dimension. Long's examples show why the dimension cannot be ignored~\cite{Long}. Here the question is more precise. For integers $n\geq1$ and $d\geq0$, let $\G(n,d)$ mean that every pure complex polynomial of total degree at most $d$ in $n$ standard Gaussian variables is safe. \begin{theorem}[Dimension--degree boundary]\label{thm:boundary} \[ \boxed{\quad \G(n,d)\ \Longleftrightarrow\ n\leq2\ \text{or}\ d\leq2\ \text{or}\ (n,d)=(3,3).\quad} \] \end{theorem} The entries below describe a universal implication for all polynomials in each dimension--degree range. Individual safe polynomials also exist outside those ranges. \begin{center} \begin{tabular}{@{}lccc@{}} \toprule & $d\leq2$ & $d=3$ & $d\geq4$\\ \midrule $n=1,2$ & purity implies safety & purity implies safety & purity implies safety\\ $n=3$ & purity implies safety & purity implies safety & failures exist\\ $n\geq4$ & purity implies safety & failures exist & failures exist\\ \bottomrule \end{tabular} \end{center} The only positive case needed beyond the existing binary and quadratic theorems is a cubic in three variables. Its quantitative form is useful: \begin{theorem}[Ternary cubics]\label{thm:cubic} Let $P\in\C[X_1,X_2,X_3]$ have total degree at most three and satisfy \eqref{eq:pure}. Then, for every nonzero polynomial $Q$, \begin{equation}\label{eq:cutoff} \E[QP^m]=0\qquad\text{whenever }m>3\deg Q. \end{equation} For $Q=0$ the conclusion is immediate. No smaller constant can replace three uniformly over this class. \end{theorem} The proof gives a classification as well as a vanishing bound. Put \begin{equation}\label{eq:circularcoords} Z=\frac{X_1+iX_2}{\sqrt2},\qquad W=\frac{X_1-iX_2}{\sqrt2},\qquad T=X_3. \end{equation} Call a phase \emph{one-sided} if all its monomials have circular weight $\deg_Z-\deg_W$ of one strict sign; $T$ has weight zero. \begin{theorem}[Normal forms for pure ternary cubics] \label{thm:normalforms} A complex polynomial of degree at most three in three Gaussian variables is pure if and only if, after a complex orthogonal linear change of coordinates and multiplication by a nonzero scalar, it is one-sided or is the exceptional phase \[ P_*=W+ZT-\frac12 Z^3. \] \end{theorem} The zero polynomial is included in the one-sided case: its empty support satisfies either strict-sign condition. Thus the classification consists of one-sided support and one exceptional orbit. One-sided phases have cutoff $m>\deg Q$; the exceptional orbit accounts for the sharp coefficient three. \subsection*{Notation and reading guide} The coordinate and parameter conventions used throughout are: \begin{center}\small \begin{tabular}{@{}p{0.22\textwidth}p{0.68\textwidth}@{}} \toprule Symbol & Meaning\\ \midrule $n,d$ & Number of real Gaussian variables and upper bound on $\deg P$\\[3pt] $Z,W,T$ & Circular pair with $\E[Z^aW^b]=\delta_{ab}a!$, and an independent standard Gaussian\\[3pt] $t$, $r$ & Formal moment parameter; $r=1/t$ near infinity, with a finite root cover when stated\\[3pt] $\Phi_Q$ & Marked series $\sum_{m\ge0}\E[QP^m]t^m/m!$\\[3pt] $q(x)$ & Gaussian exponent after a coordinate change, including its linear and constant terms\\ \bottomrule \end{tabular} \end{center} All polynomial degrees are total degrees. An expectation is the algebraic Gaussian moment functional, extended to formal series coefficient by coefficient. Complex coordinate changes preserve this functional when they preserve its bilinear covariance form. Analytic periods enter only after the comparison hypotheses are checked. For the boundary statement and its examples, read Sections~\ref{sec:introduction}--\ref{sec:cubic} and Section~\ref{sec:completion}. The full ternary-cubic proof occupies Sections~\ref{sec:formaltools}--\ref{sec:frameextraction}; the proof map below locates its two main reductions. \subsection*{Proof map} Write $P=F_3+F_2+F_1+F_0$ in homogeneous pieces. The logical route is: \begin{center}\small \begin{tabular}{@{}p{0.28\textwidth}p{0.43\textwidth}p{0.21\textwidth}@{}} \toprule Step & Statement used & Proof location\\ \midrule Find a Gaussian direction & Purity gives a nonzero $v$ with $v^Tv=0$, $\nabla F_3(v)=0$, $F_2(v)=0$ & Section~\ref{sec:frameextraction}\\[4pt] Make one coordinate affine linear & The frame test gives $P=Wh(Z,T)+G(Z,T)$ & Lemma~\ref{lem:frame}\\[4pt] Classify the affine family & One-sided support or the exceptional orbit & Section~\ref{sec:affineclassification}\\[4pt] Transfer the mixed bound & Weight cancellation or the exact shear source identity & Section~\ref{sec:shear}\\ \bottomrule \end{tabular} \end{center} Formal tools in Section~\ref{sec:formaltools} and analytic comparison in Section~\ref{sec:analyticcomparison} support the two classification steps. Section~\ref{sec:completion} assembles the conclusions. The leading cubic has nine geometric types. The abbreviated table below groups types with the same exclusion argument; the full table appears in Section~\ref{sec:frameextraction}. In every singular row, the task is to prove that purity supplies a direction passing the frame test. \begin{center}\small \begin{tabular}{@{}p{0.49\textwidth}p{0.43\textwidth}@{}} \toprule Leading cubic & Argument when the frame test fails\\ \midrule Smooth & Gradient criterion, Corollary~\ref{cor:smoothleading}\\[3pt] Nodal; transverse line--conic; nonconcurrent lines & Ordinary-node analysis, Corollary~\ref{cor:ordinarycomplete}\\[3pt] Cuspidal; tangent line--conic & Proposition~\ref{prop:cusptangentcomplete}\\[3pt] Three concurrent distinct lines & Proposition~\ref{prop:concurrentcomplete}\\[3pt] Double line plus a line; triple line & Quadratic elimination, Proposition~\ref{prop:repeatedcomplete}\\ \bottomrule \end{tabular} \end{center} Section~\ref{sec:shear} proves sharpness with a three-term example. The difficult step is reducing an arbitrary pure ternary cubic to a family with a direct moment calculation; Section~\ref{sec:cubic} states the two reductions. \subsection*{Attribution and scope} We prove the ternary-cubic normal-form classification and its sharp mixed cutoff, and deduce the dimension--degree boundary. Wilson's Theorem~1 supplies the unrestricted two-variable result and the classification by dimension alone~\cite{Wilson}. The quadratic theorem is due to Liu~\cite[Theorems~2.1 and~2.2]{Liu}; we include a short proof of the form used here. Long's cubic four-variable and quartic three-variable counterexamples are adopted from \cite[Proposition~4.1 and Theorem~5.1]{Long}, with their all-order moment identities reproduced below. For formal exponential decomposition and local Fourier--Laplace background, see Sabbah~\cite{SabbahStationary}. The polynomial stationary-frame argument below allows nonconvergent Laurent coefficients over a valued field. The analytic period comparisons are applied to rational parameter families, whose local Puiseux expansions converge. \section{A few Gaussian identities} All our generating functions are formal power series. We never need to assume that the integral of $e^{tP(X)}$ converges. Define \begin{equation}\label{eq:phi} \Phi_Q(t)=\sum_{m\geq0}\E[QP^m]\frac{t^m}{m!}. \end{equation} Then $P$ is pure exactly when $\Phi_1=1$, and is safe exactly when every $\Phi_Q$ is a polynomial. Thus a degree bound on $\Phi_Q$ gives an explicit vanishing threshold. In the circular coordinates of \eqref{eq:circularcoords}, $Z,W$ are polynomials in the independent real Gaussians $X_1,X_2$, but are not independent random variables: $W=\overline Z$. Their moments are \begin{equation}\label{eq:wick} \E[Z^aW^b]=\delta_{ab}a!. \end{equation} Rotating $(X_1,X_2)$ proves the vanishing when $a\ne b$; when $a=b$, $ZW=(X_1^2+X_2^2)/2$ has moments $a!$. An additional independent standard Gaussian $T$ has $\E[T^{2j}]=(2j-1)!!$ and zero odd moments. \begin{lemma}[One sign of weight]\label{lem:weight} Give $Z,W,T$ weights $1,-1,0$, respectively. If every monomial of $P$ has strictly positive weight, or every monomial has strictly negative weight, then $P$ is pure and \[ \E[QP^m]=0\qquad(m>\deg Q). \] \end{lemma} \begin{proof} In the positive case, every monomial of $P^m$ has weight at least $m$. A monomial of $Q$ has weight at least $-\deg Q$. For $m>\deg Q$ their product therefore has nonzero weight, and \eqref{eq:wick} kills it. The negative case follows by exchanging $Z$ and $W$. \end{proof} Complex orthogonal changes of polynomial coordinates preserve the Gaussian moment functional. This follows from its source identity $\E[e^{u^TX}]=e^{u^Tu/2}$. We use this algebraic invariance; we do not interpret a complex coordinate change as a real change of integration contour. \begin{lemma}[Linear circular contraction]\label{lem:contraction} Let $a,b$ be constants and let $F(t,z)$ be a formal power series in $t$ whose coefficients are polynomials in $z$. Then \begin{equation}\label{eq:contraction} \E\!\left[F(t,Z)e^{t(a+bZ)W}\right] =\frac{1}{1-bt}\, F\!\left(t,\frac{at}{1-bt}\right). \end{equation} \end{lemma} \begin{proof} It suffices, coefficient by coefficient in $t$, to take $F=Z^j$. Using \eqref{eq:wick}, the left side is \[ \sum_{k\geq j}t^k[z^k]z^j(a+bz)^k =\sum_{r\geq0}\binom{j+r}{j}a^j b^r t^{j+r} =\frac{(at)^j}{(1-bt)^{j+1}}. \] Every fixed coefficient uses only finitely many terms, which also justifies the statement for $t$-dependent $F$. \end{proof} \section{Quadratics in any number of variables}\label{sec:quadratics} We include a short proof of the positive quadratic result. It explains why degree two is safe in every dimension without introducing the finite-moment thresholds that are unnecessary for the boundary. The classification and sharper surrounding results are due to Liu~\cite[Theorem~2.1]{Liu}. \begin{proposition}\label{prop:quadratic} Every pure complex quadratic in $n$ standard Gaussian variables is safe. In fact \[ \E[QP^m]=0\qquad\left(m>\left\lfloor n/2\right\rfloor\deg Q\right). \] \end{proposition} \begin{proof} Write $P=X^TAX+\ell^TX+c$, with $A=A^T$, and put $\Sigma(t)=(I-2tA)^{-1}$ and $D(t)=\det(I-2tA)$. Completing the square formally gives \begin{equation}\label{eq:quad} \Phi_1(t)=D(t)^{-1/2} \exp\!\left(ct+\frac{t^2}{2}\ell^T\Sigma(t)\ell\right). \end{equation} If $\Phi_1=1$, then $D=e^S$ for a rational function $S$ regular at zero. Hence $D'/D=S'$. A rational derivative has zero residue at every finite pole, whereas $D'/D$ has the positive multiplicity as residue at every zero of $D$. Thus $D$ has no zero, so $D=1$. It follows that $A$ is nilpotent and \[ c=0,\qquad \ell^TA^j\ell=0\quad(j\geq0). \] Conversely these identities make \eqref{eq:quad} equal to one. With an additional source $u$, the same calculation now gives \begin{equation}\label{eq:quadsource} \E[e^{tP+u^TX}] =\exp\!\left(u^T\mu(t)+\tfrac12u^T\Sigma(t)u\right), \qquad \mu(t)=t\Sigma(t)\ell. \end{equation} Both $\mu$ and $\Sigma$ are polynomials in $t$. Let $s$ be the dimension of $\Span\{\ell,A\ell,A^2\ell,\ldots\}$, with $s=0$ if $\ell=0$. This space is totally isotropic, since $(A^i\ell)^TA^j\ell=\ell^TA^{i+j}\ell=0$. Nondegeneracy of the ambient form gives $s\leq h:=\lfloor n/2\rfloor$. Nilpotence gives $A^s\ell=0$, so $\deg_t\mu\leq s$. Also $\deg_t\Sigma\leq n-1\leq2h$. For a monomial of degree $q$ in $X$, a term in \eqref{eq:quadsource} uses $p$ covariance factors and $q-2p$ mean factors. Its $t$-degree is at most $2hp+h(q-2p)=hq$. Source differentiation and linearity prove the result. \end{proof} \begin{corollary}[Ternary quadratic normal form] \label{cor:quadraticnormal} Every pure quadratic in three variables has, in Gaussian coordinates, the form \[ P=Z(\alpha Z+\beta T+\gamma). \] In particular it is one-sided. \end{corollary} \begin{proof} The preceding proof gives $c=0$, nilpotent symmetric $A$, and $\ell^TA^j\ell=0$ for every $j\ge0$. If $A\ne0$, let $k$ be largest with $A^k\ne0$. In dimension three this last power has rank one, so $A^k=\lambda vv^T$ for a nonzero null vector $v$, and $Av=0$. The identity $\ell^TA^k\ell=0$ gives $v^T\ell=0$. Since $\ell$ is also null, the one-dimensional nondegenerate quotient $v^\perp/\C v$ forces $\ell\in\C v$. Choose $Z=v^TX$ and a circular partner $W$. The equation $Av=0$ makes $P$ independent of $W$, and $\tr A=0$ removes its $T^2$ coefficient. This is the displayed form. If $A=0$, the null linear form $\ell^TX$ can itself be chosen as a multiple of $Z$; the zero polynomial is immediate. \end{proof} \section{Two variables, with no degree restriction} The needed input is Wilson's support theorem~\cite[Theorem~1]{Wilson}: if a polynomial in two Gaussian variables is pure, then in the coordinates $Z,W$ every monomial has weight $\deg_Z-\deg_W$ of one strict sign. Lemma~\ref{lem:weight} immediately gives the threshold $m>\deg Q$. This proves $\G(2,d)$ for every $d$. The one-variable conclusion follows as well. Embed a one-variable polynomial into two variables by ignoring the second. A multiplier in the larger space can be averaged in that unused variable. In particular, safety in two variables implies safety in one. No extension of Wilson's theorem to three variables is being assumed here. \section{The two failures that determine the boundary}\label{sec:failures} Long's examples~\cite{Long} have the same marked generating function. Writing them side by side makes both the threshold and the failure of eventual vanishing transparent. \subsection{A quartic in three variables} Use $Z,W,T$ as above and set \[ P_3=(1+Z)\left(W-\frac{2+Z}{2}T^2\right),\qquad Q_3=Z. \] The total degree is four. For any polynomial $A$, first average over $T$, then apply Lemma~\ref{lem:contraction} with $a=b=1$. Put $z_*=t/(1-t)$. The result is \begin{align*} \E[A(Z)e^{tP_3}] &=\frac{A(z_*)}{1-t} \left(1+t(1+z_*)(2+z_*)\right)^{-1/2}\\ &=A\!\left(\frac{t}{1-t}\right), \end{align*} because the expression inside the square root is $(1-t)^{-2}$. The root with constant term one cancels the factor $(1-t)^{-1}$. \subsection{A cubic in four variables} Use two independent circular pairs $(Z_1,W_1)$ and $(Z_2,W_2)$ and set \[ P_4=(1+Z_2)\bigl(W_1(1-Z_1)+W_2\bigr),\qquad Q_4=Z_2. \] The total degree is three. For a formal parameter $s$, circular contraction gives \[ \E[e^{sW_1(1-Z_1)}]=\frac1{1+s}. \] Substitute $s=t(1+Z_2)$, then contract the second pair. Again \[ \E[A(Z_2)e^{tP_4}] =\frac{A(z_*)}{1-t}\frac{1}{1+t(1+z_*)} =A\!\left(\frac{t}{1-t}\right). \] In either example, $A=1$ gives purity, while $A=z$ gives \[ \Phi_Q(t)=\frac{t}{1-t},\qquad \E[QP_j^m]=m!\ne0\quad(m\geq1). \] Thus one linear multiplier sees a nonzero moment at every positive power. These examples disprove $\G(3,4)$ and $\G(4,3)$. \section{Why the cubic cutoff is three}\label{sec:shear} The factor three in Theorem~\ref{thm:cubic} has a concrete source. Consider \begin{equation}\label{eq:shearfamily} P=aW+(c+dZ)T-\frac{Z(c+dZ)^2}{2a},\qquad a\ne0. \end{equation} For sources $\alpha,\beta,\gamma$, first contract $W$ and then average over $T$. If $u(z)=c+dz$, the exact identity is \begin{equation}\label{eq:shearsource} \begin{split} \E[e^{tP+\alpha Z+\beta W+\gamma T}] =\exp\biggl(&\alpha(at+\beta)+\frac{\gamma^2}{2} +\gamma t\,u(at+\beta)\\[-2pt] &-\frac{\beta t}{2a}u(at+\beta)^2\biggr). \end{split} \end{equation} Setting the sources to zero proves purity. Give the sources $\alpha,\beta,\gamma$ weights $1,3,2$, respectively. Every term in the exponent has $t$-degree at most its source weight. The same is true after exponentiation. It follows that \[ \deg_t\Phi_{Z^iW^jT^k}\leq i+3j+2k\leq3(i+j+k). \] Thus every member of \eqref{eq:shearfamily} satisfies \eqref{eq:cutoff}. Now take the simplest member, \[ P_*=W+ZT-\frac12 Z^3. \] With only the $W$ source left, \eqref{eq:shearsource} becomes \[ \E[e^{tP_*+\beta W}] =\exp\!\left(-\frac{\beta t(t+\beta)^2}{2}\right). \] The highest $t$-degree in its $\beta^D$ coefficient comes only from $(-\beta t^3/2)^D/D!$. Therefore, for every $D\geq1$, \begin{equation}\label{eq:sharp} \E[W^D P_*^{3D}]=(3D)!\left(-\frac12\right)^D\ne0. \end{equation} Since $\deg W^D=D$, any proposed uniform coefficient smaller than three would already fail on this family. This establishes the sharpness part of Theorem~\ref{thm:cubic} independently of the general reduction. \begin{lemma}[The exceptional orbit]\label{lem:exceptionalorbit} If $d\ne0$, every phase in \eqref{eq:shearfamily} is equivalent to $P_*$ under a complex orthogonal linear change and nonzero scalar multiplication. If $d=0$, it is one-sided in suitable Gaussian coordinates. \end{lemma} \begin{proof} Apply the null rotation \[ Z=Z',\qquad T=T'+\lambda Z',\qquad W=W'-\lambda T'-\lambda^2Z'/2,\qquad \lambda=c/a. \] It preserves $2ZW+T^2$ and changes the phase to $aW'+dZ'T'-d^2Z'^3/(2a)$. For $d\ne0$, choose $\kappa^2=a/d$, set $Z'=\kappa Z''$, $W'=W''/\kappa$, and multiply by $\kappa/a$. The result is $P_*$. When $d=0$, the original phase is a pure linear form, so Corollary~\ref{cor:quadraticnormal} applies. \end{proof} \section{The two reductions for a ternary cubic}\label{sec:cubic} To prove the ternary-cubic theorem, we must handle a polynomial with no preferred coordinate. Write it in homogeneous pieces as \[ P=F_3+F_2+F_1+F_0. \] We want a direction in which $P$ is only affine linear. The direction must also be isotropic for the Gaussian bilinear form; an arbitrary linear change would lose the moment identities. \begin{lemma}[The algebraic frame test]\label{lem:frame} There are Gaussian coordinates $(Z,W,T)$ in which $P$ is affine linear in $W$ if and only if there is a nonzero vector $v\in\C^3$ such that \begin{equation}\label{eq:frame} v^Tv=0,\qquad \nabla F_3(v)=0,\qquad F_2(v)=0. \end{equation} \end{lemma} \begin{proof} Polarization gives \[ D_v^2P(x)=2\nabla F_3(v)^Tx+2F_2(v). \] The last two conditions in \eqref{eq:frame} say exactly that this polynomial vanishes. Complete the nonzero null vector $v$ to a Witt basis: choose a null partner with pairing one, and a unit vector in their orthogonal complement. In these coordinates $D_v$ is the $W$ derivative. Hence $D_v^2P=0$ says that no second or higher power of $W$ occurs. The converse reads off $v$ from the $W$ direction. \end{proof} The hard reduction is an existence statement, not this algebraic test. \begin{proposition}[The required Gaussian direction]\label{prop:frameextract} If a pure ternary polynomial has degree exactly three, then a vector satisfying \eqref{eq:frame} exists. \end{proposition} The proof divides according to the projective cubic $F_3=0$. Over $\C$, the list is short: a smooth cubic, an irreducible nodal or cuspidal cubic, a line with a smooth conic, three distinct lines, or a repeated line. The first group has no leading singular direction at all. In the singular cases the issue is whether the singular direction is null and whether $F_2$ vanishes there. Cases failing this test are excluded by their period behavior. Once a direction passes the test, Lemma~\ref{lem:frame} supplies the desired coordinates. Section~\ref{sec:frameextraction} proves this statement for all nine leading types. Its formal frames, regular-rank accounting and uniform cut estimates include critical degeneracies and contributions at infinity. \begin{proposition}[Classification with an affine-linear coordinate]\label{prop:wlinear} Let \[ P=W(a+bZ+cT+dZ^2+eZT+fT^2)+G(Z,T),\qquad\deg G\leq3. \] If $P$ is pure, then, up to a complex orthogonal linear change of coordinates and a nonzero scalar multiple, it is one-sided or equal to $P_*=W+ZT-Z^3/2$. In particular it is safe, with $\E[QP^m]=0$ whenever $m>3\deg Q$. \end{proposition} The proof below reduces the pure cases to the one-sign argument, the quadratic calculation, and the shear mechanism of Section~\ref{sec:shear}. Some excluded cases first reduce to a one-variable polynomial exponential period. The proof must retain the determinant factor from eliminating a Gaussian variable; omitting it can turn a decaying period into an apparent constant. It must also use the actual period selected by the Gaussian moment functional, rather than every solution of an enlarged differential equation. The next four sections separate the tools from the two reductions: formal arguments in Section~\ref{sec:formaltools}, the affine-coordinate classification in Section~\ref{sec:affineclassification}, analytic comparison in Section~\ref{sec:analyticcomparison}, and geometric frame extraction in Section~\ref{sec:frameextraction}. \section{Formal tools}\label{sec:formaltools} The first tools are algebraic: an obstruction for a cubic Gaussian series, a stationary-ball criterion over a valued field, and the corresponding criterion with logarithmic volume. They are used in both reductions. The analytic interpretation is developed separately in Section~\ref{sec:analyticcomparison}. Here a \emph{differential module} over a parameter field $K$, such as $\C(t)$ or $\C((r))$, is a finite-dimensional $K$-vector space $M$ with a connection $\nabla$ satisfying $\nabla(am)=a'm+a\nabla m$. A $K$-linear functional $J:M\to L$ into a differential extension field is \emph{horizontal} if $J(\nabla m)=(J(m))'$. The distinguished amplitude class $[1]$ is called the \emph{seed}. In a Gaussian application $J([Q])=\Phi_Q$, so purity means $J([1])=1$. The obstructions below exclude a horizontal quotient from the submodule generated by this seed to the scalar module $K$, with connection given by ordinary differentiation, taking $[1]$ to one. Lemma~\ref{lem:seedfunctional} explains why purity would supply such a quotient. \input{formal_lemmas.tex} \input{polynomial_stationary_balls.tex} \input{punctured_polynomial_frames.tex} \section{Classification with an affine-linear coordinate} \label{sec:affineclassification} We now prove Proposition~\ref{prop:wlinear}. The coefficient of $W$ determines a disjoint set of cases. Most are settled formally; the logarithmic case uses the comparison and decay argument of Section~\ref{sec:analyticcomparison}. \input{affine_coordinate_split.tex} \subsection{A complete algebraic subcase}\label{sec:algebraic} Suppose \[ P=W h(Z)+v(Z)T^2+u(Z)T+r(Z), \] where the coefficient degrees make $P$ cubic. Thus $\deg h\leq2$, $\deg v\leq1$, $\deg u\leq2$, $\deg r\leq3$. If $g=th(g)$ and $J=1-th'(g)$, formal circular contraction followed by the quadratic Gaussian identity gives \begin{equation}\label{eq:subcase} \Phi_1= \frac{\exp\!\left(tr(g)+t^2u(g)^2/(2K)\right)}{J\sqrt K}, \qquad K=1-2tv(g). \end{equation} The circular identity here is the one-variable formal Lagrange formula; it also follows directly from coefficient extraction. For completeness, for any numerator $A$, its formula is \[ \E[A(Z)e^{tWh(Z)}]=\frac{A(g)}{1-th'(g)}. \] Indeed the left side is $\sum_{m\geq0}t^m[z^m]A(z)h(z)^m$; the same expression is the formal residue of $A(z)/(z-th(z))$, whose unique small pole is $g$. The identity remains valid for the $t$-dependent numerator in \eqref{eq:subcase}. If $h(0)\ne0$, change variables formally from $t$ to $z=g(t)$ and put \[ A=h(z),\quad B=h(z)-2zv(z),\quad C=h(z)-zh'(z). \] The equation $\Phi_1=1$, after squaring, equates the exponential of a rational function with a rational function. The residue argument used in Section~\ref{sec:quadratics} forces \begin{equation}\label{eq:subcaseidentities} C^2B=A^3,\qquad 2Br+zu^2=0. \end{equation} These identities first force $\deg h\leq1$. To see this, if the distinct nonzero roots of $h$ are $\alpha_i$ with multiplicities $m_i$, then \[ \frac Ch=1-z\sum_i\frac{m_i}{z-\alpha_i}. \] When $\deg h\geq2$, the reduced numerator has the same positive degree as its denominator, has leading coefficient $1-\deg h$, and is nonzero at each $\alpha_i$. It therefore has a root away from the roots of $h$, contradicting $C^2\mid h^3$. If $h=a\ne0$, \eqref{eq:subcaseidentities} gives $v=0$ and $r=-zu^2/(2a)$. The cubic degree bound forces $\deg u\leq1$, giving exactly \eqref{eq:shearfamily}. If $h=a+bz$ with $ab\ne0$, it gives \[ v=\frac{h-h^3/a^2}{2z}. \] This has a nonzero quadratic coefficient, contradicting $\deg v\leq1$. Thus this second branch cannot be a cubic. If $h(0)=0$, the fixed point is $g=0$. Writing $b=h'(0)$, $v_0=v(0)$, $u_0=u(0)$ and $r_0=r(0)$, purity in \eqref{eq:subcase} forces \[ (1-bt)^2(1-2v_0t)=1,\qquad tr_0+\frac{t^2u_0^2}{2(1-2v_0t)}=0. \] Hence $b=v_0=u_0=r_0=0$. Every surviving monomial has strictly positive $Z$-minus-$W$ weight, so Lemma~\ref{lem:weight} applies. This proves the classification on the displayed subcase. The following arguments remove $T^3$, handle a linear $T$ coefficient of $W$, and settle the rank-one and mixed quadratic coefficients. The nondegenerate logarithmic branch is excluded in Section~\ref{sec:analyticcomparison}. \input{affine_height_cases.tex} \input{rank_one_affine_case.tex} \input{mixed_formal_exclusion.tex} \input{affine_classification_proof.tex} \section{Analytic comparison and logarithmic exclusion} \label{sec:analyticcomparison} This section supplies the period interpretation and the nondegenerate logarithmic exclusion used above. Comparison identifies the actual period; the estimates retain its normalization, full rank and behavior at every end. \input{seed_period_comparison.tex} \input{analytic_core.tex} \section{Extracting a Gaussian affine-linear frame} \label{sec:frameextraction} We prove Proposition~\ref{prop:frameextract} by considering the leading projective cubic. Each case either provides the Gaussian direction in \eqref{eq:frame} or excludes purity. The final table records the entire case split, including lower-degree terms and singular critical sets. \input{frame_extraction_cases.tex} \section{The ternary-cubic theorem and the boundary}\label{sec:completion} \subsection{The sharp mixed cutoff} Both Proposition~\ref{prop:wlinear} and the frame-extraction statement, Proposition~\ref{prop:frameextract}, are now proved. For $\deg P\leq2$, Proposition~\ref{prop:quadratic} gives the result. For degree three, frame extraction and Lemma~\ref{lem:frame} put $P$ in the family of Proposition~\ref{prop:wlinear}. A complex orthogonal linear substitution preserves both total degree and Gaussian moments. The bound therefore transfers back to every original multiplier $Q$. Equation~\eqref{eq:sharp} proves optimality, completing the proof of Theorem~\ref{thm:cubic}. \subsection{The normal-form conclusion} \begin{proof}[Proof of Theorem~\ref{thm:normalforms}] For degree at most two, Corollary~\ref{cor:quadraticnormal} gives one-sided support. For degree three, Proposition~\ref{prop:frameextract} and Lemma~\ref{lem:frame} provide an affine-linear Gaussian coordinate. Proposition~\ref{prop:wlinear} then gives one-sided support or the exceptional orbit represented by $P_*$. Conversely, one-sided phases are pure by Lemma~\ref{lem:weight}, and $P_*$ is pure by \eqref{eq:shearsource}. Gaussian linear changes and nonzero scalar multiples preserve purity. \end{proof} \subsection{The dimension--degree deduction} \begin{proof}[Proof of Theorem~\ref{thm:boundary}] Wilson proves the assertion for $n\leq2$ in every degree. Proposition~\ref{prop:quadratic} proves it for $d\leq2$ in every dimension. Theorem~\ref{thm:cubic} supplies $(n,d)=(3,3)$. Conversely, suppose none of these conditions holds. Then $n\geq3$ and $d\geq3$. If $n=3$, the excluded pair $(3,3)$ leaves $d\geq4$, and $P_3$ from Section~\ref{sec:failures} is a counterexample. If $n\geq4$, use $P_4$. Additional independent Gaussian variables can be ignored, and increasing the allowed degree does not remove an existing example. These possibilities exhaust the complement. \end{proof} The two minimal failures occur at degree four in three variables and degree three in four variables. Between them, the ternary-cubic bound is exact: a multiplier of degree $q$ can expose a nonzero moment at power $3q$, but never beyond it. \section*{Scope and further questions} \phantomsection\addcontentsline{toc}{section}{Scope and further questions} The boundary theorem leaves open the classification of individual safe phases outside its universal range, and the general converse from safety to purity. Its proof combines classical comparison and decomposition results with the explicit compact-piece, neck and tail estimates above. \clearpage \begingroup\small \phantomsection\addcontentsline{toc}{section}{References} \paragraph{AI assistance and review.} ChatGPT (OpenAI) assisted with mathematical arguments, literature checks, exposition, proof review and verification code. \begin{thebibliography}{99} \bibitem{DVEZ} H.~Derksen, A.~van den Essen and W.~Zhao, \emph{The Gaussian Moments Conjecture and the Jacobian Conjecture}, \emph{Israel Journal of Mathematics} \textbf{219} (2017), no.~2, 917--928; \href{https://arxiv.org/abs/1506.05192}{arXiv:1506.05192}. \bibitem{DLMF} NIST Digital Library of Mathematical Functions, \href{https://dlmf.nist.gov/10.25.E1}{10.25.1} and \href{https://dlmf.nist.gov/10.31.E2}{10.31.2}, modified Bessel equation and logarithmic power series. \bibitem{DrusvyatskiyLewis} D.~Drusvyatskiy and A.~S.~Lewis, \emph{Semi-algebraic functions have small subdifferentials}, \emph{Mathematical Programming, Series~B} \textbf{140} (2013), 5--29, Theorems~2.25 and~2.32; \href{https://doi.org/10.1007/s10107-012-0624-x}{doi:10.1007/s10107-012-0624-x}. \bibitem{HLTV} R.~M.~Hardt, P.~Lambrechts, V.~Turchin and I.~Voli\'c, \emph{Real homotopy theory of semi-algebraic sets}, \href{https://arxiv.org/abs/0806.0476v3}{arXiv:0806.0476v3}, Sections~2.1 and~3, especially Theorem~3.5. \bibitem{SimonGMT} L.~Simon, \emph{Introduction to Geometric Measure Theory}, author's 2017/2018 notes, Chapter~6, Section~3, Theorem~3.15, printed page~175; \href{https://math.stanford.edu/~lms/ntu-gmt-text.pdf}{author's primary text}. \bibitem{HienRoucairol} M.~Hien and C.~Roucairol, \emph{Integral representations for solutions of exponential Gau\ss--Manin systems}, \emph{Bulletin de la Soci\'et\'e Math\'ematique de France} \textbf{136} (2008), no.~4, 505--532, \href{https://doi.org/10.24033/bsmf.2564}{doi:10.24033/bsmf.2564}. \bibitem{Liu} Z.~Liu, \emph{Sharp finite Gaussian moments for quadratic polynomials}, version~3, public preprint (2026), \href{https://doi.org/10.5281/zenodo.21688251}{doi:10.5281/zenodo.21688251}. \bibitem{Long} C.~D.~Long, \emph{Small Counterexamples to the Gaussian Moments Conjecture}, public preprint (2026), \href{https://arxiv.org/abs/2607.18186v1}{arXiv:2607.18186v1}. \bibitem{Pham} F.~Pham, \emph{La descente des cols par les onglets de Lefschetz, avec vues sur Gauss--Manin}, \emph{Ast\'erisque} \textbf{130} (1985), 11--47, \href{https://numdam.org/item/AST_1985__130__11_0/}{Numdam primary text}. \bibitem{RoucairolFormal} C.~Roucairol, \emph{Formal structure of direct image of holonomic $\mathcal D$-modules of exponential type}, \emph{Manuscripta Mathematica} \textbf{124} (2007), no.~3, 299--318; \href{https://arxiv.org/abs/math/0604134v2}{arXiv:math/0604134v2}. \bibitem{SabbahStationary} C.~Sabbah, \emph{An explicit stationary phase formula for the local formal Fourier--Laplace transform}, in \emph{Singularities I}, Contemporary Mathematics \textbf{474}, American Mathematical Society, 2008, 309--330; \href{https://arxiv.org/abs/0706.3570v2}{arXiv:0706.3570v2}. \bibitem{MannaaCoquand} B.~Mannaa and T.~Coquand, \emph{Dynamic Newton--Puiseux Theorem}, public preprint (2013), Corollary~1.10, \href{https://arxiv.org/abs/1304.6770v2}{arXiv:1304.6770v2}. \bibitem{Wilson} M.~Wilson, \emph{A face-isolation proof of the two-variable Gaussian Moments Conjecture}, public preprint (2026), \href{https://arxiv.org/abs/2607.23887v1}{arXiv:2607.23887v1}. \end{thebibliography} \endgroup \end{document}