A lattice is a discrete additive subgroup of $\newcommand{\R}{\mathbb{R}}\R^n$; for the purposes of this post we’ll restrict ourselves to *full rank* lattices, i.e., those which span the entire vector space $\R^n$. An alternate way of defining a lattice is to consider it as the *integer span* of a collection of linearly independent vectors $\newcommand{\b}{\mathbf{b}}\newcommand{\c}{{,}~}\b_1\c\dotsc\c\b_n\in\R^n$, known as a *basis* of $L$. That is, $L$ consists of the collection of points

\[ \newcommand{\norm}[1]{\lVert#1\rVert}\newcommand{\x}{\mathbf{x}}\newcommand{\Z}{\mathbb{Z}} \biggl\{\,\sum_{i=1}^n x_i\b_i : x_i\in\Z \,\biggr\} . \]

The goal of this post is to show that these two ways of defining a lattice are equivalent. It is more or less immediate that a lattice $L$ in the second sense is an additive subgroup which spans $\R^n$; it is also discrete since one can show that^{1}

\[ \newcommand{\0}{\mathbf{0}} \norm{\x} \geq \min_{1\leq i\leq n}\norm{\b_i^*} \]

for all $\x\in L$ and $\x\neq\0$, where $\b_i^*$ is the Gram–Schmidt orthogonalization of $\b_i$. Thus $\0$ is an isolated point of $L$, and it follows that every point of $L$ must be isolated; a nonisolated point would imply (using a suitable translation) that $\0$ is nonisolated.

The harder direction is to show that a lattice in the first sense is also a lattice in the second sense; the remainder of the post is devoted to this. Let $L$ be a discrete additive subgroup of $\R^n$ containing $n$ linearly independent vectors.

In the case $n:=1$, since $L$ spans $\R$ it must contain $\pm a\neq0$. Since it also contains $0$ and is discrete, there must be a minimal $b>0$ with $b\in L$. Then as required we have $L= \{\, xb : x\in\Z \,\}$; if there was any point $c=xb\in L$ for $x\notin\Z$ then $c-\lfloor x\rfloor b$ would lie in $L$ and $(0,b)$, a contradiction to the minimality of $b$.

Now suppose the result holds for all lattices in $\R^{n-1}$; we will show the result holds for $L$ and appeal to induction. We know that $L$ contains $n$ linearly independent vectors; select any $n-1$ of them and let $S$ be the subspace they generate. Let $S’$ be the rotation of $S$ such that every vector in $S’$ has a $0$ in its final coordinate. Crucially, applying the rotation on $L$ to form $L’$ preserves the discrete additive subgroup structure, so we can apply the induction hypothesis to the discrete additive subgroup formed by only considering the first $n-1$ coordinates of $S’\cap L’$. Let $\b_1\c\dotsc\c\b_{n-1}$ be a basis of this lattice (which exists by hypothesis), except extended with an extra $0$ coordinate so as to live in $\R^n$ instead of $\R^{n-1}$.

Since $L’$ generates $\R^n$, it must contain a vector not in $S’$, i.e., with nonzero final coordinate. Furthermore, it must contain a vector with *minimal* positive final coordinate; otherwise there would exist a sequence $\newcommand{\N}{\mathbb{N}}\{\x_i\in L’\}_{i\in\N}$ with $(x_i)_n\to0$ as $i\to\infty$. By translating the $\x_i$ by suitable multiples of $\b_1\c\dotsc\c\b_{n-1}$ we can ensure that they lie in the compact set

\[ \biggl\{\, \sum_{i=1}^{n-1} \alpha_i \b_i + (0,\dotsc,0,\alpha) : \alpha_i\in[0,1]\c\lvert \alpha\rvert\leq\max_{i\in\N}\lvert(x_i)_n\rvert \,\biggr\} , \]

and therefore some subsequence of the $\x_i$ converges to some point in the set. Taking successive differences of this subsequence we get a sequence of points in $L’$ which converge to $\0\in L’$, a contradiction to the discreteness of $L’$.

Let $\b_n\in L’$ be a vector with minimal nonzero final coordinate. We claim that $\b_1\c\dotsc\c\b_n$ is a basis for $L’$. Let $\x\in L’$ be arbitrary, and consider

\[ \x’ := \x – \biggl\lfloor \frac{x_n}{(b_n)_n} \biggr\rfloor \b_n \in L’ . \]

Its final coordinate is $x_n-\lfloor x_n/(b_n)_n\rfloor(b_n)_n\in[0,(b_n)_n)$ and therefore by minimality of $(b_n)_n$ it must be $0$, so $\x’\in S’\cap L’$ and therefore can be written as an integer combination of $\b_1\c\dotsc\c\b_{n-1}$. Thus $\x=\x’+\lfloor x_n/(b_n)_n\rfloor\b_n$ can be written as an integer linear combination of $\b_1\c\dotsc\c\b_n$, and so these vectors form a basis of $L’$. Applying the reverse rotation of $L\mapsto L’$ to the basis $\b_1\c\dotsc\c\b_n$ gives us a basis of $L$, as required.

- To see this, write $\x = \sum_{i=1}^n x_i \b_i = \sum_{i=1}^n x_i^* \b_i^*$. Let $k$ be the largest $i$ such that $x_i\neq0$, so $\DeclareMathOperator{\sp}{span} \x \in x_k \b_k^* + \sp(\b_1,\dotsc,\b_{k-1})$ which shows that $x_k^*=x_k\in\Z$. Then
\begin{equation}\norm{\x}^2 = \sum_{i=1}^k (x_i^*)^2 \norm{\b_i^*}^2 \geq (x_k^*)^2 \norm{\b_k^*}^2 \geq \min_{1\leq i\leq n} \norm{\b_i^*}^2 , \end{equation}

as required. ↩