Let us consider two vectors $\mathbf{a}=(a_1,...,a_n)$ and $\mathbf{b}=(b_1,...,b_m)$ so that each quantity is an integer $a_i,b_j \in \mathbb{N}$. It represents for example supply and demand. Let $\mathbf{C}=(c_{ij})_{ij} \in \mathbb{R}_{+}^{n \times m}$ be a matrix representing the cost of transporting goods from $i$ to $j$.

My question is the following: can we guarantee that a solution of the Optimal transport problem between $\mathbf{a}$ and $\mathbf{b}$ is a matrix whose elements are integer?

More precisely I am interested in the solutions of the following Optimal Transport problem: \begin{equation} \underset{\boldsymbol\pi \in \Pi(\mathbf{a},\mathbf{b})}{\min} \sum_{i=1,j=1}^{n,m}c_{ij}\pi_{ij}=\underset{\pi \in \Pi(\mathbf{a},\mathbf{b})}{\min} \langle \mathbf{C},\boldsymbol\pi \rangle \end{equation}

where $\Pi(\mathbf{a},\mathbf{b})$ is the "transport" polytope defined by: \begin{equation} \Pi(\mathbf{a},\mathbf{b})=\{\boldsymbol\pi \in \mathbb{R}^{n \times m} \ | \pi_{ij}\geq 0; \ \forall j, \ \sum_{i=1}^{n} \pi_{ij}=b_j; \ \forall i, \ \sum_{j=1}^{m} \pi_{ij}=a_i\} \end{equation} In this way $\pi_{ij}$ is the amount of goods transferred from $i$ to $j$ so that the overall amount is preserved. (Remark: usually this problem is considered with histograms but here I am not normalizing the quantities.)

We can solve this problem using the simplex algorithm for example.

More formally my question is: can we guarantee that there exists an optimal transport plan $\boldsymbol\pi^{*}$ of the previous problem such that $\pi^{*}_{ij}$ is an integer?

  • My hint:

The problem is a standard linear program (LP) as it can be written: \begin{equation} \underset{\begin{smallmatrix}\mathbf{p} \in \mathbb{R}_{+}^{nm} \\ \mathbf{A}\mathbf{p}=\begin{pmatrix} \mathbf{a} \\ \mathbf{b} \end{pmatrix} \end{smallmatrix}}{\min} \mathbf{c}^{\top} \mathbf{p} \end{equation}

by vectorizing all the quantities, i.e. $\mathbf{c}$ is $nm$ vectors equal to the stacked columns contains in $\mathbf{C}$ and same for $\mathbf{p}$ which corresponds to $\boldsymbol\pi$. In this form the constraint matrix writes: \begin{equation} \mathbf{A}=\begin{pmatrix}\mathbf{1}_n^{\top} \otimes I_m \\ I_n \otimes \mathbf{1}_m^{\top} \end{pmatrix} \in \mathbb{R}^{(n+m)\times nm} \end{equation}

where $\otimes$ is the Kronecker product, $I_n$ the identity matrix of size $n$ and $\mathbf{1}_n$ is a vector of ones of size $n$. This problem is a relaxation of the Integer Linear Programming (ILP) problem where elements of $\mathbf{p}$ are constrained to be integers (which is want we want). Solving (ILP) is NP-hard in general and we can use the (LP) relaxation and round the entries but the solution may not be optimal, and even not feasible.

However as written in [1] (see also https://en.wikipedia.org/wiki/Integer_programming) elements of an optimal solution of the (LP) are integers when $\mathbf{A}$ is totally unimodular, that is every square submatrix of $\mathbf{A}$ has determinant $0,+1$ or $-1$. Is $\mathbf{A}$ totally unimodular in this case?

[1] Alexander Schrijver, Theory of Linear and Integer Programming, John Wiley & Sons, 1986


Yes, this is true and can found, for example in Chapter 7 of the "Handbook of Discrete and Computational Geometry" edited by Csaba D. Toth, Joseph O'Rourke, Jacob E. Goodman (chapter by Alexander Barvinok).

  • $\begingroup$ Thank you very much! $\endgroup$ Oct 26 '20 at 12:29
  • $\begingroup$ More generally, network flow problems have totally unimodular constraint matrices. A sufficient condition is to have two nonzero entries in each column, with exactly one $1$ and exactly one $-1$. $\endgroup$
    – RobPratt
    Oct 26 '20 at 14:20

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.