PureOne's picture
Publish verified v5.0.0 research release
b6f1c7f verified
Raw History Blame Contribute Delete
7.06 kB
\section{The intrinsic and extrinsic geometries}\label{sec:geometry}
The finite-grid energy depends on the individual integer characters. Its infinitesimal limit combines them into a positive quadratic form. Keeping these two objects separate is the central geometric discipline of the paper.
\begin{theorem}[Metric, chords, and normal curvature]\label{thm:geometry}
The map $F_{A,w}$ is a smooth embedding of $\T^d$ when $A$ is primitive. Its induced Riemannian metric is the constant matrix
\begin{equation}\label{eq:metric}
G=A^\top W A=\sum_iw_i a_i a_i^\top,\qquad W=\diag(w).
\end{equation}
For arbitrary phases $\theta,\eta$,
\begin{equation}\label{eq:chord}
\norm{F(\theta)-F(\eta)}^2
=4\sum_iw_i\sin^2\!\left(\frac{a_i\cdot(\theta-\eta)}2\right).
\end{equation}
The intrinsic curvature vanishes. The extrinsic second fundamental form, however, is generally nonzero and equals
\begin{equation}\label{eq:II}
\II_\theta(h,k)_i=-\sqrt{w_i}e^{ia_i\cdot\theta}(a_i\cdot h)(a_i\cdot k).
\end{equation}
In particular,
\begin{equation}\label{eq:quarticII}
\norm{\II_\theta(k,k)}^2=\sum_iw_i(a_i\cdot k)^4.
\end{equation}
\end{theorem}
\begin{proof}
The $i$th coordinate derivative in direction $h$ is $i\sqrt{w_i}e^{ia_i\cdot\theta}(a_i\cdot h)$. Taking the real inner product of derivatives in directions $h,k$ gives $\sum_iw_i(a_i\cdot h)(a_i\cdot k)=h^\top Gk$. Positivity of weights and full rank make $G$ positive definite, so the map is an immersion. It is injective by Lemma~\ref{lem:faithful}; an injective immersion of a compact manifold into Euclidean space is an embedding. Formula~\eqref{eq:chord} follows from $|e^{iu}-e^{iv}|^2=4\sin^2((u-v)/2)$.
The metric coefficients are constant in the lifted phase coordinates, hence their derivatives, the Christoffel symbols, and the intrinsic Riemann curvature tensor are all zero. Differentiating a second time gives the right side of \eqref{eq:II}. Its real inner product with any tangent vector is zero: the product in every coordinate is a purely imaginary number times real factors. Thus the entire second derivative is normal, and it is the second fundamental form because the domain connection in these coordinates is zero. Taking its squared norm proves \eqref{eq:quarticII}.
\end{proof}
\subsection{The lattice and the intrinsic invariants}
For every primitive $A$, not necessarily unimodular, define
\begin{equation}\label{eq:shortest}
s(A,w)=\min_{k\in\Z^d\setminus\{0\}}k^\top Gk.
\end{equation}
Positive definiteness ensures that the minimum is positive and attained on a finite set. The torus is $\R^d/(2\pi\Z^d)$ with metric $G$, so
\begin{align}
\Vol(\T^d,G)&=(2\pi)^d\sqrt{\det G},\label{eq:volume}\\
\sys(\T^d,G)&=2\pi\sqrt{s(A,w)},\label{eq:sys}\\
\inj(\T^d,G)&=\pi\sqrt{s(A,w)}.\label{eq:inj}
\end{align}
These are standard flat-torus identities. To see them directly, a fundamental coordinate cube has side $2\pi$ and volume density $\sqrt{\det G}$. Every nontrivial deck translation is $2\pi k$, whose shortest representative is the straight segment of length $2\pi\sqrt{k^\top Gk}$. The shortest nontrivial loop minimizes over nonzero $k$, and the injectivity radius is half the shortest deck displacement. More generally,
\begin{equation}\label{eq:geodesic}
d_G(\theta,\eta)=\min_{k\in\Z^d}\sqrt{(\theta-\eta+2\pi k)^\top G(\theta-\eta+2\pi k)}.
\end{equation}
This intrinsic geodesic distance is not the Euclidean chord in \eqref{eq:chord}.
\begin{theorem}[Unimodular systole and its equality cases]\label{thm:systole}
If $A$ is unimodular, then $s(A,w)=\lambda_A(w)$. The shortest nonzero integer vectors are exactly $\pm z_D$ for minimum-weight cocircuits $D$. In particular there are $2N_*(A,w)$ such vectors.
\end{theorem}
\begin{proof}
For a nonzero $k\in\Z^d$, the support of $Ak$ contains a cocircuit. Every nonzero integer coordinate has square at least one, so
\[
k^\top Gk=\sum_iw_i(Ak)_i^2
\ge\sum_{i\in\supp(Ak)}w_i\ge\lambda_A(w).
\]
Lemma~\ref{lem:unitcocircuit} realizes equality for each minimum cocircuit. If equality holds, positivity forces the support to be exactly one such cocircuit, and each of its nonzero integer coordinates to have magnitude one. The complement hyperplane then forces $k$ to be a scalar multiple of its primitive normal $z_D$; an integer multiple other than $\pm1$ would have larger energy. This also proves the count.
\end{proof}
The minimum-cut version for graph lattices is not historically new; it is related to shortest-vector computations in lattices of Voronoi's first kind \cite{mckilliam}. Regular-matroid lattice and code connections also predate this manuscript \cite{gurjar}. Here they are presented with full proofs as consequences of a common spectral character model.
\subsection{Heat trace and the minimum-cocircuit multiplicity}
Let $\Delta_G$ be the positive Laplace--Beltrami operator on the intrinsic flat torus. This is an infinite-dimensional differential operator, distinct from the finite magnetic matrix used later. Its eigenfunctions $e^{ik\cdot\theta}$ have eigenvalues $k^\top G^{-1}k$, so its heat trace is
\begin{equation}\label{eq:theta}
\Theta_G(t)=\sum_{k\in\Z^d}e^{-t k^\top G^{-1}k},\qquad t>0.
\end{equation}
Poisson summation gives
\begin{equation}\label{eq:poisson}
\Theta_G(t)=\frac{\pi^{d/2}\sqrt{\det G}}{t^{d/2}}
\sum_{\ell\in\Z^d}e^{-\pi^2\ell^\top G\ell/t}.
\end{equation}
For completeness, with Fourier transform $\widehat f(\xi)=\int_{\R^d}f(x)e^{-2\pi ix\cdot\xi}\,dx$, the Gaussian $f(x)=e^{-t x^\top G^{-1}x}$ has transform $\pi^{d/2}\sqrt{\det G}\,t^{-d/2}e^{-\pi^2\xi^\top G\xi/t}$. Summing this identity over the dual integer lattices yields \eqref{eq:poisson}; the sums converge absolutely.
\begin{corollary}[First heat correction]\label{cor:heat}
For unimodular $A$, put
\[
R_G(t)=\frac{t^{d/2}}{\pi^{d/2}\sqrt{\det G}}\Theta_G(t)-1.
\]
Then
\begin{equation}\label{eq:heatleading}
R_G(t)=2N_*(A,w)e^{-\pi^2\lambda_A(w)/t}(1+o(1))\quad(t\downarrow0).
\end{equation}
Thus $-t\log R_G(t)/\pi^2$ tends to $\lambda_A(w)$, and $\tfrac12e^{\pi^2\lambda_A(w)/t}R_G(t)$ tends to $N_*(A,w)$.
\end{corollary}
\begin{proof}
In the nonzero dual-lattice sum in \eqref{eq:poisson}, Theorem~\ref{thm:systole} identifies the least energy and exactly $2N_*$ minimizers. There is a strictly larger next energy because a bounded ellipsoid contains finitely many lattice points. To justify neglecting the infinite remainder, fix $t_0>0$ and the positive gap $\delta$ above the least energy. For $t<t_0$, factor $e^{-\pi^2\lambda_A(w)/t}$ out of the remainder and dominate its residual exponential series by a constant times $e^{-\pi^2\delta(1/t-1/t_0)}$. The constant is finite by Gaussian convergence at $t_0$. The ratio of the remainder to the leading term tends to zero.
\end{proof}
Recovering the first heat correction at small $t$ is exponentially ill-conditioned as a numerical measurement. Equation~\eqref{eq:heatleading} is a structural identity, not an efficient min-cut algorithm or a proposal to replace combinatorial algorithms by delicate subtraction of heat traces.