\subsection{One exclusion criterion for several geometric cases} \label{sec:gradientexclusion} The smooth case does not need a list of critical-point formulas. The same argument also treats many singular leading forms. Its hypotheses must hold in every complex direction. A bound only on the original real integration space would not suffice. We state the criterion for a polynomial exponential amplitude, then specialize it to the Gaussian. Norms are Hermitian norms on complex affine space; the Gaussian quadratic remains holomorphic. \begin{theorem}[A gradient criterion]\label{thm:gradientcriterion} Let $P,q\in\C[x_1,\ldots,x_n]$, with $P$ nonconstant and $\deg q\leq J$, where $J\geq1$. Suppose that outside a compact set \begin{equation}\label{eq:gradientconditions} \|\nabla P(x)\|\geq c\|x\|^{J-1},\qquad \frac{\|\operatorname{Hess}P(x)\|}{\|\nabla P(x)\|^2}\leq C \end{equation} for some $c>0$ and $C<\infty$. For every fixed polynomial $A$, no nonzero flat period of \[ A(x)e^{tP(x)-q(x)}\,dx_1\cdots dx_n \] is a finite exponential-polynomial sum $\sum_\lambda e^{\lambda t}p_\lambda(t)$ with polynomial $p_\lambda$. In particular it cannot be a nonzero constant. \end{theorem} \begin{proof} We first justify the full cycle space and the tails. The gradient is nonzero outside a compact set, so the affine critical set is compact and algebraic, hence finite. The gradient bound is stronger than the condition excluding vanishing cycles at infinity. The finite-critical thimble decomposition therefore applies, with the whole local Milnor homology at each isolated critical point, not just its Morse approximation~\cite[Part~I, Sections~1--4]{Pham}. When several critical points have the same value, take disjoint Milnor balls and apply excision to their complement. This gives the direct sum of their full local homology before any values are grouped. Here the Gaussian or polynomial amplitude can be kept on these fixed $P$-cycles. Choose a generic argument $\theta$ of $t$. Outside the local critical neighbourhoods, lift a descending phase ray by \[ P(x)=\lambda-e^{-i\theta}s,\qquad \frac{dx}{ds}=-e^{-i\theta} \frac{\overline{\nabla P(x)}}{\|\nabla P(x)\|^2}. \] For large $s$ the lift is outside the compact set, because its phase value is unbounded. The first condition in \eqref{eq:gradientconditions} gives $\|x\|^J\leq C_1(1+s)$. The real differential of this vector field has norm at most a constant times the second ratio in \eqref{eq:gradientconditions}, hence is bounded. Transported $(n-1)$-cycle volumes grow at most as $\exp(C_2s)$; the radial speed is bounded as well. Also $|A(x)e^{-q(x)}|\leq C_3(1+s)^M\exp(C_4s)$ for some $M$. Thus $\exp(-|t|s)$ dominates the complete tail, uniformly for large $|t|$ in a smaller sector. The same estimate applies to every fixed polynomial amplitude. We must also check that these are a full basis for the actual phase, rather than selected admissible cycles. Put \[ F_\sigma=P-\sigma q/t,\qquad 0\leq\sigma\leq1. \] Since $\|\nabla q\|\leq C_5\|\nabla P\|$ outside the same compact set, all large $|t|$ give \[ \|\nabla F_\sigma\|\geq\tfrac12\|\nabla P\|. \] The ratio involving the Hessian stays uniformly bounded too. Indeed $\|\operatorname{Hess}q\|=O(\|x\|^{J-2})$ for $J\geq2$, while that Hessian is zero for $J=1$. All critical points and their values therefore remain in fixed compact sets. On the preimage of a phase half-plane beyond those values, the parameter field \[ W_\sigma=\frac{q}{t} \frac{\overline{\nabla F_\sigma}}{\|\nabla F_\sigma\|^2} \] satisfies $\partial_\sigma F_\sigma+dF_\sigma(W_\sigma)=0$ and $\|W_\sigma\|\leq C_6(1+\|x\|)/|t|$. Its flow is complete for $0\leq\sigma\leq1$. Extend it by a cutoff away from that half-plane. It gives an isotopy of the ambient relative pairs defining the full rapid-decay homology. This explicitly prevents an infinity class from appearing during the deformation. The fixed $P$-thimbles have uniformly descending tails for every $F_\sigma$, by the preceding estimates. Truncate them beyond a common large phase value. Their boundaries stay in the descending half-plane throughout the isotopy, so they define its flat relative classes. They are a full basis at $\sigma=0$ and remain one at $\sigma=1$. Finite critical points may merge, and their critical values may coincide; the argument uses the full relative pairs and does not require a Morse family. The same cycles are flat as $t$ varies in the chosen sector. These improper chains and their homotopies are admissible for the period pairing. One can integrate first on truncated chains and then pass to the limit: the tail bound makes every boundary term tend to zero. Indeed the differential of $W_\sigma$ grows at most as $C(1+\|x\|^J)/|t|$; along the controlled tails this is $C'(1+s)/|t|$. The homotopy volumes have at most this exponential growth, still dominated by the descending factor for all large $|t|$. Thus Stokes and differentiation under the integral hold throughout; the relative-pair calculation does not depend on a convergence claim for the original real Gaussian integral. It remains to determine the local asymptotics at a finite critical point with value $\lambda$. For an analytic amplitude germ, the vanishing-cycle density is a finite sum of terms $s^{\alpha+k}(\log s)^j$, with analytic coefficient series and $\alpha>-1$. Its Laplace asymptotics consequently have the form \[ e^{\lambda t}\sum_{\alpha,k,j} c_{\alpha,k,j}\,t^{-\alpha-k-1}(\log t)^j, \qquad k\geq0. \] This is the local analytic result of Pham~\cite[Part~II, Section~1.1, (1.5)--(1.6) and the footnote on printed page~26]{Pham}; it permits degenerate isolated critical points. Pham credits Maisonobe and Rombaldi for the local solution comparison used there. The amplitude $A e^{-q}$ is analytic. Every nonzero term has a strictly negative power of $t$ and tends to zero after removing $e^{\lambda t}$. The tails outside the local neighbourhood are exponentially negligible by the uniform bound already proved. The finite-critical decomposition accounts for the full module. Group equal critical values before separating the distinct formal exponential factors. Within each such factor the displayed local expansions contain only decaying powers and logarithms. A nonzero polynomial prefactor cannot occur. The zero function is allowed when all these contributions vanish. This proves the theorem. \end{proof} \begin{corollary}[Original Gaussian markers]\label{cor:gradientmarkers} For the standard Gaussian quadratic $q=\sum x_i^2/2$, suppose \eqref{eq:gradientconditions} holds with $J=2$. Then every original marked series $\Phi_A$ is either zero or is not a finite exponential-polynomial sum. In particular $P$ is neither pure nor safe. \end{corollary} \begin{proof} The exact top de Rham quotient and connection are \[ M=\C(t)[x]\Big/\sum_i(\partial_i+tP_i-x_i)\C(t)[x], \qquad\nabla[A]=[A_t+PA]. \] Gaussian integration by parts makes $[A]\mapsto\Phi_A$ horizontal on this vector-space quotient. The spatial and parameter derivatives commute; no ordinary Jacobian-ideal quotient is substituted. If some $\Phi_A$ is a nonzero finite exponential-polynomial sum, a constant-coefficient differential operator isolates one exponential and kills its polynomial prefactor, giving $c e^{\beta t}$, $c\ne0$. Because \[ \partial_t\Phi_A=\Phi_{PA}, \] this is still an original fixed polynomial marker. Replace $P$ by $P-\beta$ and divide the marker by $c$. Its formal seed value is one. Lemma~\ref{lem:seedfunctional} gives an actual constant period for that marker, contradicting Theorem~\ref{thm:gradientcriterion}. Its comparison hypotheses hold with base $t$ and phase $tP-q$ on $\mathbb A^1\times\mathbb A^n$, after removing the finite exceptional base set. No convergence of the original Gaussian series is assumed. Finally $\Phi_1(0)=1$, so it is a nonzero series; purity or safety would make it a polynomial. \end{proof} \begin{corollary}[Smooth leading form, in every degree] \label{cor:smoothleading} If $\deg P=D\geq2$ and the homogeneous leading form satisfies $\nabla F_D(v)=0$ only for $v=0$, then $P$ is not pure. In fact Corollary~\ref{cor:gradientmarkers} holds for every original marker. \end{corollary} \begin{proof} Compactness of the complex unit sphere gives $\|\nabla P(x)\|\geq c\|x\|^{D-1}$ for large $\|x\|$. Also $\|\operatorname{Hess}P\|=O(\|x\|^{D-2})$, so its ratio to the squared gradient is $O(\|x\|^{-D})$. Both conditions for $J=2$ follow, with all lower terms retained. \end{proof} \begin{lemma}[The ordinary-node test]\label{lem:nodecoercivity} Let $P=F_3+F_2+F_1+F_0$ in $n\geq2$ variables. Suppose its projective singular set is finite and every singular point $[v]$ has $\ker\operatorname{Hess}F_3(v)=\C v$. If $F_2(v)\ne0$ at every such point, then $\|\nabla P(x)\|\geq c\|x\|$ outside a compact set. Consequently $P$ is not pure under the Gaussian functional. \end{lemma} \begin{proof} Otherwise choose an unbounded sequence with $\|\nabla P(x)\|/\|x\|\longrightarrow0$. A projective limit $[v]$ is singular for $F_3$. Choose a fixed complement to $\C v$ and write $x=sv+y$, where $y=o(s)$. With $H=\operatorname{Hess}F_3(v)$, homogeneity gives the exact expansion \[ \nabla P(sv+y)=sHy+\nabla F_3(y)+s\nabla F_2(v) +\nabla F_2(y)+\nabla F_1. \] The map $H$ is injective on the chosen complement. Divide by $s$. The error from $\nabla F_3(y)/s$ is $o(\|y\|)$, so the equation first forces $y$ to stay bounded. A limit $y_0$ satisfies $Hy_0+\nabla F_2(v)=0$. Pair bilinearly with $v$. Symmetry and $Hv=0$ kill the first term, whereas Euler's identity gives $v\cdot\nabla F_2(v)=2F_2(v)$, a contradiction. The Hessian of a cubic is $O(1+\|x\|)$, so the second condition in \eqref{eq:gradientconditions} follows from the first. Apply Corollary~\ref{cor:gradientmarkers}. \end{proof} \subsection{A common residual family for ordinary nodes} \label{sec:ordinaryresidual} For a pure cubic with only ordinary leading nodes, the preceding lemma forces $F_2$ to vanish at some node. An isotropic such node gives the frame already proved in Lemma~\ref{lem:frame}. Every nonisotropic such node has the following single normal form. \begin{lemma}\label{lem:ordinarynormal} Let $[w]$ be an ordinary node of a ternary $F_3$, with $F_2(w)=0$ and $w^Tw\ne0$. An invertible affine change of coordinates puts the pair consisting of $P$ and the Gaussian quadratic into \begin{align*} P={}&T(uv+h)+a u^3+d v^3+b u^2+c v^2+e u+f v+k,\\ q={}&\tfrac12(Au^2+Bv^2+mT^2+2Cuv+2DuT+2EvT) +\ell_u u+\ell_v v+\ell_T T+q_0, \end{align*} where $m\ne0$ and $\det\operatorname{Hess}q\ne0$. All Gaussian means and cross terms must be retained. Its leading curve is irreducible nodal when $ad\ne0$, a transverse line and smooth conic when exactly one of $a,d$ vanishes, and three nonconcurrent lines when $a=d=0$. \end{lemma} \begin{proof} Take $T$ in the direction $w$ and choose a complementary plane. The singular-point equations make $F_3$ linear in $T$; its coefficient is a nondegenerate binary quadratic because the node is ordinary. Factor that quadratic and choose its two linear factors as coordinates $u,v$. Thus $F_3=Tuv+B_3(u,v)$. Replace $T$ by $T$ plus a linear form in $u,v$ to remove the $u^2v$ and $uv^2$ terms of $B_3$. There is no $T^2$ term in $F_2$, precisely because $F_2(w)=0$. The full coefficient of $T$ is therefore $uv+\alpha u+\beta v+h_0$. Translations of $u,v$ make it $uv+h$. The cubic terms outside $T$ remain pure cubes. A constant shift of $T$ removes the remaining quadratic $uv$ term. This gives the displayed $P$. The direction of $T$ has not changed, so $m\ne0$; invertibility of the affine change preserves the nondegenerate Gaussian Hessian, while translations and the shifts of $T$ produce the means and cross terms shown. If $ad\ne0$, the polynomial $Tuv+a u^3+d v^3$ is irreducible: an independent factor of a polynomial linear in $T$ would divide both $uv$ and $a u^3+d v^3$, whose gcd is one. Its gradient shows that its sole singular point is the stated ordinary node. If $a=0,d\ne0$, it is $v(Tu+d v^2)$, with a smooth conic meeting the line transversely at two points; the other one-zero case is symmetric. If both vanish, it is the coordinate triangle $Tuv$. \end{proof} This consolidates three leading-curve rows into one residual family. It is a reduction, not yet an exclusion of that family. In particular the full quadratic $q$, including its means, cannot be replaced by a convenient diagonal form during the remaining period argument. \subsection{Two controls on the scope of the criterion} Ordinary tameness of $P$ alone is insufficient for the Gaussian conclusion. Let $P=\ell^Tx$, with $\ell=(1,i)/\sqrt2$ in two variables. Its gradient is constant and nonzero, but $\Phi_1=\exp(t^2\ell^T\ell/2)=1$. The phase $tP-q$ has its critical point at $x=t\ell$, escaping with action zero. Thus an isolated critical set for $P$, or the weaker gradient bound away from zero, would lose exactly the contribution that purity needs. Condition~\eqref{eq:gradientconditions} with $J=2$ prevents this. The distinction between an original marker and a time-dependent amplitude also remains essential. For $P=(X_1^2+X_2^2)/2$ one has $\Phi_1=(1-t)^{-1}$ and $(1-t)\Phi_1=1$. This phase satisfies the gradient criterion. The unit in its rational differential module is represented by a time-dependent amplitude. No original fixed polynomial marker has a nonzero polynomial period, by Corollary~\ref{cor:gradientmarkers}. The abstract module alone does not record that distinction.