8
$\begingroup$

For $n > 0$, let $P$ be an integral polytope, that is, the convex hull in $\mathbb{R}^n$ of points in $\mathbb{Z}^n$. Suppose that $\dim(P) = n$.

Question: Given $d > n + 2$ is it true that

$$ dP \cap \mathbb{Z}^n \stackrel{?}{=} (n+2)P \cap \mathbb{Z}^n + (d-(n+2))P \cap \mathbb{Z}^n \ ? $$

That is for any integer point $u$ in the dilation $dP$ is there an integer point $v$ in the dilation $(n+2)P$ such that $u-v$ is an integer point of the dilation $(d-(n+2))P$? If so, is the construction explicit? If not are there conditions on $P$ where it does hold?

$\endgroup$
7
  • 1
    $\begingroup$ This is closely related to the integer decomposition property (IDP), or being "integrally closed" or being "a normal polytope". $\endgroup$ Commented Apr 28, 2015 at 20:32
  • $\begingroup$ is this not trivially obvious? $\endgroup$
    – JMP
    Commented Apr 29, 2015 at 4:43
  • $\begingroup$ @PerAlexandersson I completely agree. If $P$ has IDP or is integrally closed then the decomposition is possible. Are there conditions when this happens? Is there a construction for the decomposition? What if it doesn't have IDP? I've looked here and they prove that $kP$ will have IDP for all $k \geq n - 1$, but I don't think that's enough to prove this. They also provide an example where $2P\cap\mathbb{Z}^n \neq P\cap\mathbb{Z}^n + P\cap\mathbb{Z}^n$, but for $P$ with $\dim(P) \geq 7$. Is there an easy counterexample in my case? $\endgroup$
    – hypercube
    Commented Apr 29, 2015 at 16:54
  • $\begingroup$ @JonMarkPerry Is it trivially obvious that it holds or doesn't? Can you elaborate a little more please? $\endgroup$
    – hypercube
    Commented Apr 29, 2015 at 16:56
  • 2
    $\begingroup$ @JonMarkPerry: that the RHS is contained in the LHS is obvious, but it is not obvious that every point in the LHS decomposes as required in the RHS. Consider the following example. Let $P$ be the regular simplex inscribed in the unit cube, with vertices (0,0,0), (1,1,0), (1,0,1) and (0,1,1). The point $(1,1,1)$ is in $2P$ but it cannot be decomposed as the sum of two integer points in $P$. This is usually expressed as $P$ is not integrally closed or $P$ does not have the IDP property. $\endgroup$ Commented May 16, 2015 at 19:08

2 Answers 2

7
$\begingroup$

The answer is YES in the following stronger form:

Lemma 1: Let $d>n$ and let $p\in dP\cap \mathbb{Z}^n$. Then there is an $i\in\{1,\dots, n\}$ and points $p'\in iP\cap \mathbb{Z}^n$, $p_1,\dots, p_{d-i}\in P\cap \mathbb{Z}^n$ such that $$ p = p' + p_1 + \cdots + p_{d-i}. $$

For the proof I will (as is customary in this type of problems) work on the cone over $P$, defined as follows: $$ C_P := \operatorname{pos}(P\times \{1\}) \subset \mathbb{R}^{n+1}. $$ Observe that $dP$ can be identified with $C_P \cap \{x_{n+1}=d\}$. I call height of a point in $C_P$ the value of its last coordinate, so that integer points at height $d$ are identified with integer points in $dP$. Lemma 1 is equivalent to:

Lemma 2: Let $p\in C_P\cap \mathbb{Z}^{n+1}$ be an integer point at height $d> n$. Then, there is a point $p'\in C_P$ at a certain height $i\in \{1,\dots,n\}$ and integer points points $p_1,\dots, p_{d-i}$ such that $p=p' + p_1 + \cdots + p_{d-i}$.

Proof of Lemma 2: There is no loss of generality in assuming that $P$ is a simplex (if not, consider an integer triangulation of $P$ and apply the result to each individual simplex). So, let $v_1,\dots,v_{n+1}$ be the vertices of $P$, considered as points in $C_P$ at height one, and let $Z$ be the unit parallelepiped obtained as the Minkowski sum of the segments $Ov_i$.

The cone $C_P$ can be tiled by integer translations of $Z$; more precisely, it is tiled by all translations $v+Z$ where $v$ runs over the integer non-negative combinations of the $v_i$'s. Thus, $p$ can be decomposed as $p'+v$, where $p'$ is an integer point in $Z$ and $v$ is an integer non-negative combination of the $v_i$'s.

Except in the trivial cases where $p'=O$ or $p'$ is the sum of all the $v_i$'s, the integer point $p'$ in $Z$ has height between $1$ and $n$, which finishes the proof. QED

$\endgroup$
4
  • $\begingroup$ Thank you for the answer! I'm glad that the result holds! I have a few questions though if you don't mind: 1. What does pos mean in the definition of the cone? 2. Can you explain why the cone is tiled by integer translations of $Z$? 3. How constructive is this process? If I gave you $p$ could you give an algorithm to determine $p'$ and the $p_i$'s? $\endgroup$
    – hypercube
    Commented May 17, 2015 at 15:30
  • $\begingroup$ 1. "pos" means "positive span"; that is, if $A$ is a subset of a real vector space, $pos(A)$ is the cone of linear combinations of elements of $A$ with nonnegative coefficients. $\endgroup$ Commented May 17, 2015 at 20:56
  • $\begingroup$ 2. By a change of basis (take $v_1,\dots,v_n$ as your basis) you can think of $C_P$ as the positive orthant and of $Z$ as the unit cube $[0,1]^n$. The positive orthant is tiled by integer-translated copies of the unit cube. $\endgroup$ Commented May 17, 2015 at 20:57
  • $\begingroup$ 3. A more constructive proof of the same Lemma is as follows. As before, suppose $P$ is a simplex. Any point $p$ in $dP$ can be written as a (perhaps not integral) combination $\lambda_1 v_1 + \cdots + \lambda_{n+1} v_{n+1}$ of the vertices of $P$, with $\sum_i \lambda_i=d$ and all $\lambda_i\ge 0$. Since $d\ge n+1$, some $\lambda_i$ is $\ge 1$. That means that $p - v_i$ is in $(d-1)P$. Iterate until you get a point at height $\le n$. $\endgroup$ Commented May 17, 2015 at 21:05
3
$\begingroup$

Somewhat belatedly (four years later), I noticed that the Shapley-Folkman lemma (https://en.wikipedia.org/wiki/Shapley%E2%80%93Folkman_lemma) yields a strong version of the result. SF says (as a special case) that if $S$ is a subset of $n$-dimensional Euclidean space, and we set $P$ to be its convex hull, then when $d > n$, for all $x$ in $dP$, there exists $y $ in $nP$ such that $x - y$ is a sum of $n-d$ elements of $S$.

So let $S = P \cap {\bf Z^n}$. For $x$ in $dP \cap {\bf Z^n}$, SF implies there is a decomposition $x = y + w$ where $w$ is a sum of $d-n$ elements of $S$ and $y \in dP$. Then $y = x-w$ is in ${\bf Z^n}$ (since $x$ and $w$ are). Hence $y \in nP \cap {\bf Z^n}$, yielding a stronger result: $dP \cap {\bf Z^n} = nP \cap {\bf Z^n} + ( S + \dots + S)$ with $d-n$ copies of $S$.

$\endgroup$

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .