Lattice coverings by parallelotopes containing cube vertices

Write the vertices of Q_3 as ijk, where i,j,k\in\{0,1\}, and set A=000, B=100, C=110, D=010, A_1=001, B_1=101, C_1=111, and D_1=011. Up to symmetries of the cube, there are three configurations of five vertices:

(i) A,B,C,D,A_1; (ii) A,B,D,A_1,C_1; (iii) A,B,C,D_1,C_1.

(i) The covering follows from Theorem 2.

(ii) We triangulate the prism ABCA_1B_1C_1 using the edges AC_1, AB_1, B_1C. The prism ADCA_1D_1C_1 is triangulated by translating AB_1 and B_1C onto the opposite facets.

(iii) We triangulate the prism ABCA_1B_1C_1 using the edges AC_1, BA_1, BC_1. The prism ADCA_1D_1C_1 is triangulated by translating BA_1 and BC_1 onto the opposite facets.

In both cases, (ii) and (iii), the triangulations induced on opposite facets of Q_3 agree under translation. The resulting triangulation therefore extends periodically to \Bbb R^3 and has vertex set \Bbb Z^3. Moreover, the difference of the endpoints of every edge is, up to sign, the difference of two of the five selected vertices. Hence the covering follows from Lemmas 5 and 4.

Reflecting the first n-1 coordinates, if necessary, we may assume that \mathbf{1}:=e_1+\cdots+e_n\in P in place of e_n\in P. These reflections preserve both Q_{n-1}\times\{0\} and the integer lattice. For any permutation \tau of \{1,2,\dots,n\}, consider the simplex

Note that \mathcal{T}:=\{T_{\tau}: \tau\in S_n\} is a triangulation of Q_n that can be extended by translations to a triangulation of \Bbb R^n with vertex set \Bbb Z^n. For any two vertices u,v of T_{\tau}, we have

If n\notin A, the sum is a vertex of Q_{n-1}\times\{0\}; if n\in A, it is the difference between \mathbf{1} and another vertex of that facet. Since 0,\mathbf{1}\in P, in both cases u-v\in P-P. Hence Lemma 5 shows that each simplex can be translated into P. The periodic triangulation has uniformly bounded simplex diameters, so Lemma 4 gives P+\Bbb Z^n=\Bbb R^n.

Let p be the apex of the considered pyramid. Write the vertices of its base as a+U\varepsilon, \varepsilon\in\{0,1\}^{n-1}, where U=(u_1,\dots,u_{n-1}) is an n\times(n-1) matrix. Then the volume of the pyramid is |\det(U,p-a)|/n. Thus it is enough to prove

Let r_j=(u_{j1},\dots,u_{j,n-1}) be the j-th row of U, and put

Since every vertex a+U\varepsilon lies in Q_n, the range of the j-th coordinate has length P_j+N_j. Hence

Moreover,

Let c_j be the cofactor of the j-th entry in the last column of (U,v), so that

By reflecting suitable coordinates of Q_n and, if necessary, changing the common sign of the cofactors, we may assume c_j\ge0 for all j. From 0\le p_j\le1 and (?),

and

It therefore remains to show that both sums on the right are at most 1.

Consider first \sum_jc_j(1-N_j). Form the n\times n matrix Y whose j-th row is y_j=(r_j,1-N_j). Expansion in the last column gives

By (?), each row has the convex decomposition

After multiplying the last coordinate by -1, the vectors appearing in this decomposition become reduced node–arc incidence vectors of a directed graph. The node–arc incidence matrix of a directed graph is totally unimodular; see, e.g., [S, p. 274]. Hence every square matrix formed from these vectors has determinant 0 or \pm1. Since each y_j is a convex combination of these vectors, multilinearity of the determinant gives |\det Y|\le1. Therefore

Replacing U by -U interchanges P_j and N_j and changes all cofactors by the same sign. The same argument therefore gives

This proves (Proof).

Let

where the f_i are linear functionals (corresponding to the opposite facets) and a_i,\Delta_i \in \Bbb R, \Delta_i>0, 1\le i\le n. It follows that

Let \sigma_0,\ldots,\sigma_n be the vertices of \sigma. Suppose that \sigma_i-\sigma_j\in P-P for every i,j, and set

Thus, M_i-m_i\le \Delta_i. Hence, for every \sigma_j, 0\le j\le n, we have

Since P is nondegenerate, the functionals f_i are linearly independent, so there exists t\in\Bbb R^n with f_i(t)=m_i-a_i. It follows that \sigma\subset t+P. The converse is immediate.